{
 "title": "Agent Audit Readiness 2027 — public documentation scorecard",
 "publisher": "whiteboxx.ai",
 "checkedAt": "2026-11-20",
 "statusValues": {
  "documented": "Clear official documentation found",
  "partial": "Limited, preview/beta, or covers only part of the criterion",
  "not found": "No official public documentation found; not a claim that the capability is absent"
 },
 "criteria": [
  {
   "id": "traces",
   "label": "Run/trace logs",
   "definition": "Per-run record of steps, tool calls and model calls available to customers."
  },
  {
   "id": "citations",
   "label": "Source attribution",
   "definition": "Outputs cite the data sources or documents used."
  },
  {
   "id": "lineage",
   "label": "Data lineage/versioning",
   "definition": "Records which version or retrieval time of source data was used in a run (agent/configuration versioning alone does not qualify)."
  },
  {
   "id": "approvals",
   "label": "Human approval controls",
   "definition": "Configurable human approval before actions / human-in-the-loop."
  },
  {
   "id": "scoping",
   "label": "Permission scoping",
   "definition": "Per-agent limits on tools, data and actions (least privilege)."
  },
  {
   "id": "budgets",
   "label": "Budget/usage limits",
   "definition": "Enforced budget or usage limits per agent or per run (spend, credits, tokens, iterations); org-wide caps only count as partial."
  },
  {
   "id": "auditExport",
   "label": "Audit log export/API",
   "definition": "Admin audit logs exportable to a SIEM or available via API."
  },
  {
   "id": "retention",
   "label": "Retention controls",
   "definition": "Configurable retention of logs and transcripts."
  },
  {
   "id": "evals",
   "label": "Evaluation/testing tools",
   "definition": "Built-in evaluation tooling or test sets."
  }
 ],
 "platforms": [
  {
   "name": "Microsoft Copilot Studio",
   "vendor": "Microsoft",
   "productUrl": "https://learn.microsoft.com/en-us/microsoft-copilot-studio/",
   "scopeNote": "Assessed Microsoft Copilot Studio (standard harness) agents and agent flows, including the Microsoft Purview, Dataverse and Power Platform admin center controls that the Copilot Studio docs point to; checked 2026-11-20.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://learn.microsoft.com/en-us/microsoft-copilot-studio/authoring-review-activity",
     "note": "The Activity page records each activity as an activity map with transcript, showing per-node inputs, outputs, tool/action calls, knowledge queries and errors; it applies only to agents with generative orchestration, and historical activity requires an Exchange mailbox."
    },
    "citations": {
     "status": "documented",
     "url": "https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio",
     "note": "Knowledge-grounded chat responses can include in-text citations linking to the knowledge source used, though citations are not shown in voice conversations or added automatically to customized responses."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Searched Copilot Studio and Purview docs; the activity map and Purview AccessedResources list which sources were used, but no documentation was found recording the version or retrieval time of source data per run."
    },
    "approvals": {
     "status": "documented",
     "url": "https://learn.microsoft.com/en-us/microsoft-copilot-studio/flows-request-for-information",
     "note": "Agent flows support a Request for information human-review action that pauses execution until a designated reviewer responds, and multistage human/AI approvals (labeled preview) are documented separately."
    },
    "scoping": {
     "status": "documented",
     "url": "https://learn.microsoft.com/en-us/microsoft-copilot-studio/security-and-governance",
     "note": "Admins can govern knowledge sources, connectors, actions, HTTP requests and triggers through data policies, tools can run with the end user's credentials, and with Agent 365 agents can be governed as Entra identities with Conditional Access and role-based controls."
    },
    "budgets": {
     "status": "documented",
     "url": "https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management",
     "note": "Admins can set a monthly Copilot Credit consumption limit for an individual agent in the Power Platform admin center, but the docs note that limit does not stop agent flows the agent calls."
    },
    "auditExport": {
     "status": "documented",
     "url": "https://learn.microsoft.com/en-us/purview/ai-copilot-studio",
     "note": "Copilot Studio prompts, responses and maker activities are captured in the Purview unified audit log, which is searchable in Purview, ingestible into Microsoft Sentinel, and exposed through the Office 365 Management API CopilotInteraction schema."
    },
    "retention": {
     "status": "documented",
     "url": "https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-transcript-controls",
     "note": "Admins can turn transcript saving to Dataverse on or off per environment or environment group and change the Dataverse transcript retention period, and Purview retention policies can retain or delete agent interactions."
    },
    "evals": {
     "status": "documented",
     "url": "https://learn.microsoft.com/en-us/microsoft-copilot-studio/analytics-agent-evaluation-intro",
     "note": "Agent evaluation lets makers generate, import or write test sets, run them with multiple test methods including safety evaluators, and trigger runs via REST API or connectors for CI/CD."
    }
   }
  },
  {
   "name": "Salesforce Agentforce",
   "vendor": "Salesforce",
   "productUrl": "https://www.salesforce.com/agentforce/",
   "scopeNote": "Assessed Agentforce agents (Agent Builder / Agentforce Studio, Agent Script) with Agentforce Session Tracing, the Einstein Trust Layer audit data in Data 360 and Testing Center; help.salesforce.com pages could not be fetched directly, so several findings rely on official search-result snippets and are scored conservatively.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://help.salesforce.com/s/articleView?language=en_US&id=ai.generative_ai_session_trace_data_model.htm&type=5",
     "note": "The Agentforce Session Tracing data model stores session, interaction and interaction-step records in Data 360, including references to each LLM call the reasoning engine makes."
    },
    "citations": {
     "status": "documented",
     "url": "https://developer.salesforce.com/docs/ai/agentforce/guide/citations.html",
     "note": "Agents can cite sources for responses, via the Show Sources option on Agentforce Data Libraries, citation settings on prompt template actions, or Apex citation classes (platform-generated citations only for agents created after May 26, 2025)."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Agent/configuration versioning is documented, but no documentation was found recording the version or retrieval time of source data used in a run; agent versions and Data 360 retriever versions are documented, but this is configuration versioning; no documentation was found recording the source data version or retrieval time per run."
    },
    "approvals": {
     "status": "documented",
     "url": "https://developer.salesforce.com/docs/ai/agentforce/guide/ascript-ref-actions.html",
     "note": "Actions can be set to require user confirmation before the agent executes them (require_user_confirmation in Agent Script, Require user confirmation in the builder), recommended for actions that change records or send email."
    },
    "scoping": {
     "status": "documented",
     "url": "https://help.salesforce.com/s/articleView?language=en_US&id=ai.agent_user.htm&type=5",
     "note": "Customer-facing agents run as a dedicated agent user whose permission sets, role and sharing determine what data and actions the agent can access, with guided setup creating a minimal-access agent user."
    },
    "budgets": {
     "status": "partial",
     "url": "https://help.salesforce.com/s/articleView?language=en_US&id=ai.agent_reasoning_engine.htm&type=5",
     "note": "The docs state a fixed limit of up to seven reasoning loops per request for legacy-builder agents, but no configurable per-agent spend or credit cap was found (Digital Wallet consumption tracking is org-level)."
    },
    "auditExport": {
     "status": "partial",
     "url": "https://developer.salesforce.com/docs/ai/agentforce/guide/otel-api.html",
     "note": "A Beta REST API exports a session's full trace in OpenTelemetry format from the Session Tracing and Generative AI Audit and Feedback data models, but it is per session and limited to sessions from the previous 72 hours."
    },
    "retention": {
     "status": "partial",
     "url": "https://help.salesforce.com/s/articleView?language=en_US&id=ai.generative_ai_trust_arch2.htm&type=5",
     "note": "Salesforce says audit and feedback data is stored in the customer's Data 360 instance, where the customer controls how long it is kept, but no specific retention configuration for session tracing data was documented."
    },
    "evals": {
     "status": "partial",
     "url": "https://help.salesforce.com/s/articleView?id=ai.agent_testing_center.htm&type=5&language=en_US",
     "note": "Agentforce Testing Center supports generated or uploaded batch test cases evaluated for response accuracy, topic/subagent routing, action execution and knowledge retrieval, but the Agentforce Studio test workflow is labeled Beta."
    }
   }
  },
  {
   "name": "ServiceNow AI Agents",
   "vendor": "ServiceNow",
   "productUrl": "https://www.servicenow.com/docs/r/intelligent-experiences/ai-agent-studio.html",
   "scopeNote": "Assessed ServiceNow AI agents and agentic workflows built in AI Agent Studio (installed as ServiceNow Otto AI Agents in the Brazil release docs); the official docs site is sparsely indexed and many short links returned 404, so items documented only in ServiceNow Community articles were not counted.",
   "scores": {
    "traces": {
     "status": "partial",
     "url": "https://www.servicenow.com/docs/r/intelligent-experiences/ai-agent-studio.html",
     "note": "Official docs say the AI Agent Studio Activity page holds execution logs for AI agents and agentic workflows, filterable by version, but the step and tool-call detail of those logs was confirmed only in community content."
    },
    "citations": {
     "status": "not found",
     "url": null,
     "note": "Searched ServiceNow docs for AI agent citation or source attribution; nothing official found."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Agent/configuration versioning is documented, but no documentation was found recording the version or retrieval time of source data used in a run; docs reference version control that tracks changes to instructions sent to the LLM and version filters on execution logs, which is agent versioning only, not data lineage."
    },
    "approvals": {
     "status": "partial",
     "url": "https://downloads.docs.servicenow.com/pdf/enus/servicenow-zurich-release-notes-enus.pdf",
     "note": "Zurich release notes say agentic playbooks can keep humans in the loop for verification, approval or manual steps; the per-tool Supervised execution mode was described only in community articles."
    },
    "scoping": {
     "status": "documented",
     "url": "https://www.servicenow.com/docs/r/zurich/intelligent-experiences/acls-and-rolemasking.html",
     "note": "ACLs control who can invoke an agent, the agent runs as an AI user or dynamic user, and role masking limits a dynamic-user agent to the overlap of the invoker's roles and the agent's permitted roles, applied down to tool ACLs."
    },
    "budgets": {
     "status": "not found",
     "url": null,
     "note": "Searched ServiceNow docs for per-agent usage, assist consumption or iteration limits; nothing official found."
    },
    "auditExport": {
     "status": "not found",
     "url": null,
     "note": "Searched ServiceNow docs for AI agent audit log export or SIEM integration; only community and press-release claims about AI Control Tower were found."
    },
    "retention": {
     "status": "not found",
     "url": null,
     "note": "Searched ServiceNow docs for retention settings on AI agent execution logs; nothing official found."
    },
    "evals": {
     "status": "documented",
     "url": "https://www.servicenow.com/docs/r/intelligent-experiences/execute-aia-eval.html",
     "note": "Agentic evaluation runs test an agent or workflow against a dataset of execution logs using metrics such as overall task completeness plus custom metrics (Brazil release docs)."
    }
   }
  },
  {
   "name": "Google Gemini Enterprise Agent Platform",
   "vendor": "Google",
   "productUrl": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/overview",
   "scopeNote": "Assessed Gemini Enterprise Agent Platform (the current name for Vertex AI Agent Builder / Agent Engine, now Agent Runtime) plus Gemini Enterprise (formerly Agentspace) where noted, such as Workflow Builder approvals.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/tracing",
     "note": "Deployed agents can emit OpenTelemetry traces to Cloud Trace, with spans for LLM calls and tool calls and optional prompt and response capture (the LlamaIndex integration is Preview)."
    },
    "citations": {
     "status": "documented",
     "url": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/grounding/grounding-with-vertex-ai-search",
     "note": "Grounded responses return grounding metadata (chunks and supports) that map answer segments to the source documents retrieved, which can be used to display citations."
    },
    "lineage": {
     "status": "partial",
     "url": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/reference/rest/v1beta1/SessionEvent",
     "note": "Session events (v1beta1 reference) are timestamped and can carry grounding metadata, but no documentation was found recording the version of the source data used per run."
    },
    "approvals": {
     "status": "documented",
     "url": "https://docs.cloud.google.com/gemini/enterprise/docs/workflow-builder/use-hitl-steps",
     "note": "Gemini Enterprise Workflow Builder has Request info and Approval human-in-the-loop steps that pause a run until a person responds; the page states no preview status."
    },
    "scoping": {
     "status": "documented",
     "url": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/agent-identity-overview",
     "note": "Each agent gets its own SPIFFE-based IAM principal for least-privilege grants, constrained by IAM deny and Principal Access Boundary policies, VPC Service Controls, and Agent Gateway policies on tool calls."
    },
    "budgets": {
     "status": "partial",
     "url": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/policies/token-management-overview",
     "note": "Token management policies can enforce token budgets grouped by individual agent identity, but the feature is labeled early preview."
    },
    "auditExport": {
     "status": "documented",
     "url": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/optimize/observability/overview",
     "note": "Agent audit, gateway and execution logs, with agent and user attribution, can be routed to Google Security Operations, Pub/Sub for third-party SIEMs, BigQuery or locked Cloud Storage."
    },
    "retention": {
     "status": "documented",
     "url": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/optimize/observability/overview",
     "note": "Agent logs can be routed to custom log buckets with configurable retention of 1 to 3,650 days and optional locking, and Gemini Enterprise admins can also set an assistant chat retention period."
    },
    "evals": {
     "status": "documented",
     "url": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/evaluation-agents",
     "note": "The evaluation service scores agents on final responses and tool-call trajectories using predefined and custom metrics, and the docs also describe online evaluation monitors for live traffic."
    }
   }
  },
  {
   "name": "Amazon Bedrock AgentCore",
   "vendor": "Amazon Web Services",
   "productUrl": "https://aws.amazon.com/bedrock/agentcore/",
   "scopeNote": "Assessed Amazon Bedrock AgentCore (Runtime, harness, Gateway, Policy, Memory, Observability, Evaluations), plus Amazon Bedrock Agents and Knowledge Bases where AgentCore itself has no equivalent (citations, action confirmation).",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness-operations.html",
     "note": "AgentCore Observability writes OpenTelemetry traces to CloudWatch for every harness invocation, covering model calls, tool invocations, memory operations and shell commands with timing and payloads; CloudWatch Transaction Search must be turned on first."
    },
    "citations": {
     "status": "documented",
     "url": "https://docs.aws.amazon.com/bedrock/latest/APIReference/API_agent-runtime_Citation.html",
     "note": "Bedrock Agents (InvokeAgent) and Knowledge Bases (RetrieveAndGenerate) return a citations field that links parts of the answer to the knowledge base sources retrieved; this is a Bedrock Agents/Knowledge Bases feature, not a native AgentCore Runtime output."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Agent/configuration versioning is documented, but no documentation was found recording the version or retrieval time of source data used in a run; agentCore Runtime records which runtime version served each request (aws.agentcore.runtime.version span attribute), but that is agent versioning only; no per-run record of source-data versions or retrieval times was found."
    },
    "approvals": {
     "status": "documented",
     "url": "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness-lifecycle-hooks.html",
     "note": "Harness lifecycle hooks (added September 2026) can run a synchronous Lambda before a tool call that allows or denies it, and inline function tools pause so client code can get human approval; Bedrock Agents also offers a per-function requireConfirmation setting."
    },
    "scoping": {
     "status": "documented",
     "url": "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/policy-core-concepts.html",
     "note": "AgentCore Policy evaluates Cedar policies on every Gateway tool call, by principal, tool and input parameters; harness allowedTools lists, IAM execution roles and per-user OAuth scoping through AgentCore Identity add further limits."
    },
    "budgets": {
     "status": "documented",
     "url": "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness-operations.html",
     "note": "The harness takes per-invocation hard caps (maxIterations, timeoutSeconds, maxTokens) plus session lifetime limits, set as harness defaults or overridden on a single invocation."
    },
    "auditExport": {
     "status": "partial",
     "url": "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/observability.html",
     "note": "AgentCore records API activity in AWS CloudTrail, which can be exported to a SIEM, but the docs say CloudTrail coverage currently applies only to certain capabilities (for example Gateway, Agent Registry and harness operations)."
    },
    "retention": {
     "status": "partial",
     "url": "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-create-a-memory-store.html",
     "note": "AgentCore Memory has a configurable event expiry period for stored conversation events, and telemetry goes to customer-owned CloudWatch log groups, but the AgentCore docs do not describe a retention setting for traces themselves."
    },
    "evals": {
     "status": "documented",
     "url": "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/dataset-evaluations-on-demand.html",
     "note": "AgentCore Evaluations (generally available) provides built-in and custom evaluators and runs them on demand or in batches against managed datasets with optional ground truth."
    }
   }
  },
  {
   "name": "ChatGPT Workspace Agents",
   "vendor": "OpenAI",
   "productUrl": "https://openai.com/business/workspace-agents/",
   "scopeNote": "Assessed ChatGPT workspace agents in ChatGPT Enterprise/Edu (GA May 2026) because they have public admin and security documentation; OpenAI Frontier is sold as its enterprise agent platform but its product page had no public admin docs (and refused automated fetches), and AgentKit's Agent Builder and Evals shut down on November 30, 2026.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://cdn.openai.com/business-guides-and-resources/workspace-agents-security-overview.pdf",
     "note": "OpenAI's workspace agents security overview (April 29, 2026) says the Compliance API exposes traces for every agent run, and compliance logs include run, connector-call, skill, trigger and memory events."
    },
    "citations": {
     "status": "partial",
     "url": "https://help.openai.com/en/articles/12628342-company-knowledge-in-chatgpt-business-enterprise-and-edu",
     "note": "ChatGPT company knowledge documents clickable citations to connected sources, but no official page found says workspace agent outputs include citations."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Agent/configuration versioning is documented, but no documentation was found recording the version or retrieval time of source data used in a run; builders can review an agent's version history and republish earlier versions, but that is agent versioning only; no per-run record of which data versions were used was found."
    },
    "approvals": {
     "status": "documented",
     "url": "https://cdn.openai.com/business-guides-and-resources/workspace-agents-security-overview.pdf",
     "note": "Write-capable app actions require human approval by default, and builders can set approval behavior per action (Always ask, Never ask or Custom)."
    },
    "scoping": {
     "status": "documented",
     "url": "https://cdn.openai.com/business-guides-and-resources/workspace-agents-security-overview.pdf",
     "note": "Admins control which apps and actions are available through RBAC and app settings, and builders can limit each agent to enabled tools, connection scopes and parameter constraints on its actions."
    },
    "budgets": {
     "status": "partial",
     "url": "https://cdn.openai.com/business-guides-and-resources/workspace-agents-security-overview.pdf",
     "note": "Usage limits and credit pools can be set at the workspace, role/group or user level, but the overview says budget caps or alerts for individual agents are not currently offered."
    },
    "auditExport": {
     "status": "documented",
     "url": "https://cdn.openai.com/business-guides-and-resources/workspace-agents-security-overview.pdf",
     "note": "The Compliance Logs Platform exports agent lifecycle, run and connector events as immutable JSONL files for SIEM, DLP and eDiscovery, and the Compliance API also logs every change to every agent."
    },
    "retention": {
     "status": "partial",
     "url": "https://cdn.openai.com/business-guides-and-resources/workspace-agents-security-overview.pdf",
     "note": "Agent content follows the workspace's ChatGPT retention settings, with no agent-specific retention control, and OpenAI's help center says the Compliance Logs Platform keeps data for only 30 days."
    },
    "evals": {
     "status": "not found",
     "url": null,
     "note": "Searched help.openai.com and openai.com for evaluation or test tools for workspace agents and found only preview-before-publish; the separate platform Evals product is being shut down on November 30, 2026."
    }
   }
  },
  {
   "name": "Claude Managed Agents",
   "vendor": "Anthropic",
   "productUrl": "https://claude.com/platform/claude-managed-agents",
   "scopeNote": "Assessed Claude Managed Agents (beta, managed-agents-2026-04-01 header) on the Claude Developer Platform, plus the organization-level Compliance API, rather than claude.ai Claude Enterprise chat or the self-hosted Claude Agent SDK.",
   "scores": {
    "traces": {
     "status": "partial",
     "url": "https://platform.claude.com/docs/en/managed-agents/session-observability",
     "note": "The Console session viewer (beta) shows each session's model requests, tool calls with inputs and results, and raw events, which can be downloaded as JSON; OpenTelemetry tracing and the full event history API are also documented; scored partial because Managed Agents as a whole is labeled beta."
    },
    "citations": {
     "status": "partial",
     "url": "https://platform.claude.com/docs/en/build-with-claude/citations",
     "note": "The Claude API documents citations for documents, search results and web search, but no page found says Managed Agents session outputs carry source citations."
    },
    "lineage": {
     "status": "partial",
     "url": "https://platform.claude.com/docs/en/managed-agents/memory",
     "note": "Sessions can be pinned to a specific agent version, and each memory write creates an immutable memory version linked to the session (kept about 30 days), but which source-data versions a run read is not recorded."
    },
    "approvals": {
     "status": "partial",
     "url": "https://platform.claude.com/docs/en/managed-agents/permission-policies",
     "note": "Permission policies per tool (always_allow, always_ask, auto) can pause the session for approval before a tool runs; MCP toolsets default to always_ask (beta); scored partial because Managed Agents as a whole is labeled beta."
    },
    "scoping": {
     "status": "partial",
     "url": "https://platform.claude.com/docs/en/managed-agents/tools",
     "note": "Each agent's configuration turns individual built-in tools on or off, can restrict web search and fetch to listed domains, and attaches specific MCP servers, while memory stores can be mounted read-only; scored partial because Managed Agents as a whole is labeled beta."
    },
    "budgets": {
     "status": "partial",
     "url": "https://platform.claude.com/docs/en/managed-agents/budgets",
     "note": "Session budgets (beta) set a hard list-cost ceiling per session, and the session pauses with stop reason budget_reached once it is hit; outcome loops also take max_iterations; scored partial because Managed Agents as a whole is labeled beta."
    },
    "auditExport": {
     "status": "partial",
     "url": "https://platform.claude.com/docs/en/manage-claude/compliance-activity-feed",
     "note": "The Compliance API Activity Feed provides an organization-wide audit log via API, including platform memory-store events and kept for 6 years, but the docs do not confirm it covers Managed Agents session or agent events."
    },
    "retention": {
     "status": "partial",
     "url": "https://platform.claude.com/docs/en/managed-agents/overview",
     "note": "Session transcripts are kept until you delete them through the API, Managed Agents cannot use Zero Data Retention (except when self-hosted), and no setting for automatic retention periods was found."
    },
    "evals": {
     "status": "partial",
     "url": "https://platform.claude.com/docs/en/managed-agents/define-outcomes",
     "note": "Outcomes (beta) set up a separate grader that scores a session's work against a rubric and loops until it passes, but no test-set or regression-testing tool for agents was found."
    }
   }
  },
  {
   "name": "IBM watsonx Orchestrate",
   "vendor": "IBM",
   "productUrl": "https://www.ibm.com/products/watsonx-orchestrate",
   "scopeNote": "Assessed IBM watsonx Orchestrate (SaaS on IBM Cloud and AWS) with its Agent Development Kit (ADK), agent analytics/control plane and Agent Identity, from ibm.com/docs, developer.watson-orchestrate.ibm.com and IBM's official tutorials and announcements.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://www.ibm.com/docs/en/watsonx/watson-orchestrate/base?topic=monitoring-agents",
     "note": "Agent analytics gives a Traces view per message, with span-level detail of model decisions, tool executions, knowledge retrieval and workflow actions; the page is labeled legacy experience and analytics are not supported on-premises."
    },
    "citations": {
     "status": "documented",
     "url": "https://developer.watson-orchestrate.ibm.com/knowledge_base/build_kb",
     "note": "Knowledge base configuration includes a citations_shown setting that controls how many source citations appear in answers, and search response fields can map to citation metadata."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Agent/configuration versioning is documented, but no documentation was found recording the version or retrieval time of source data used in a run; aDK release notes describe releasing and switching between multiple agent versions, but that is agent versioning only; no per-run record of knowledge-source versions was found."
    },
    "approvals": {
     "status": "partial",
     "url": "https://developer.watson-orchestrate.ibm.com/agents/agent_design/agentic_workflow_decision_guide",
     "note": "Agentic workflows can include User Activity steps that pause and route an approval task to a named person or role, but no per-tool-call approval setting for agents was found."
    },
    "scoping": {
     "status": "documented",
     "url": "https://www.ibm.com/new/announcements/agent-identity-in-ibm-watsonx-orchestrate-is-now-generally-available",
     "note": "Agent Identity (GA November 3, 2026) gives each agent its own registered identity, with OAuth tokens scoped to both the agent and the user it acts for."
    },
    "budgets": {
     "status": "partial",
     "url": "https://developer.watson-orchestrate.ibm.com/agents/build_agent",
     "note": "Native agents have a per-response max-token setting and a context compaction threshold, but no per-agent or per-run spend, credit or iteration cap was found beyond plan-level usage limits."
    },
    "auditExport": {
     "status": "documented",
     "url": "https://www.ibm.com/docs/en/watsonx/watson-orchestrate/base?topic=logs-activity-tracking-events-cloud",
     "note": "Activity tracking events go to IBM Cloud Logs and other targets through Activity Tracker Event Routing; AWS-hosted instances can export audit logs to an S3 bucket after a support request, and Agent Identity adds a per-agent access log."
    },
    "retention": {
     "status": "documented",
     "url": "https://www.ibm.com/docs/en/watsonx/watson-orchestrate/base?topic=agent-configuring-data-retention-time-ai-chat",
     "note": "Admins can set how many days chat history is kept before automatic deletion, while the monitoring docs describe a fixed 30-day retention for traces and agent analytics."
    },
    "evals": {
     "status": "documented",
     "url": "https://developer.watson-orchestrate.ibm.com/evaluate/create_data",
     "note": "The ADK evaluation framework builds test cases from recorded chats or user stories, runs them against an agent, and reports expected versus actual tool calls and journey success; evaluation REST APIs are also documented."
    }
   }
  },
  {
   "name": "SAP Joule Studio (Joule agents)",
   "vendor": "SAP",
   "productUrl": "https://www.sap.com/products/artificial-intelligence/joule-studio/custom-ai-agents.html",
   "scopeNote": "Assessed SAP Joule Studio (the new Joule Studio within Joule Work) and Joule Studio, classic edition, using help.sap.com docs; the page bodies render with JavaScript, so most evidence comes from official help.sap.com search-index text plus the Joule Development Guide.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://help.sap.com/docs/JOULESTUDIO/7e496fe6f81f4975aec07d9c1cbd2aef/d1aa69f85b7a4f7d997aef6d9be6bb95.html",
     "note": "Monitoring a Solution says each run opens a trace detail view showing the agent's reasoning and action steps and its model calls, and the CLI's jl solution traces command lists executions with span detail."
    },
    "citations": {
     "status": "partial",
     "url": "https://help.sap.com/docs/Joule_Studio_classic/45f9d2b8914b4f0ba731570ff9a85313/626498be11064953810a852535900ca7.html",
     "note": "The classic edition documents a document grounding tool for agents, but the new Joule Studio's Known Issues page says document grounding is not supported, and no general citation behavior for custom agents was found (some SAP-delivered agents, such as SuccessFactors Performance Preparation, mention inline citations)."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Searched help.sap.com for data versions or retrieval timestamps recorded per agent run and found none; only solution and agent versioning is described."
    },
    "approvals": {
     "status": "documented",
     "url": "https://help.sap.com/docs/Joule_Studio_classic/45f9d2b8914b4f0ba731570ff9a85313/e1d5bd4898654a09b82f991755108bd1.html",
     "note": "In Joule Studio, classic edition, each tool added to an agent has a Require user confirmation before running option for human-in-the-loop control; no equivalent was found in the new Joule Studio docs."
    },
    "scoping": {
     "status": "partial",
     "url": "https://help.sap.com/docs/JOULESTUDIO/7e496fe6f81f4975aec07d9c1cbd2aef/bbdfc06ed2d0408dbe5b3595e9e36284.html",
     "note": "Per-agent access policies in SAP Cloud Identity Services control which users can invoke each deployed agent, and the classic best-practices page advises limiting each agent's toolset, but no per-agent least-privilege model for data and actions was found."
    },
    "budgets": {
     "status": "documented",
     "url": "https://help.sap.com/docs/JOULE/ba88d1ec6a1b442098863d577c19b0c0/f037b144232d491caad84a7950262e26.html",
     "note": "The Joule Development Guide's agent configuration properties include an iterations setting (1 to 100) per agent, and the Joule Work desktop admin guide documents a per-user daily cap on agent actions."
    },
    "auditExport": {
     "status": "partial",
     "url": "https://help.sap.com/docs/Joule_Studio_classic/45f9d2b8914b4f0ba731570ff9a85313/ed3f72f449a743a6b38ddf6d0a8f44f7.html",
     "note": "The classic edition logs security-related events to the SAP BTP Audit Log service, and Joule Work provisioning mentions an optional SAP Audit Log Viewer with download, but no SIEM or API export specific to agents was found."
    },
    "retention": {
     "status": "partial",
     "url": "https://help.sap.com/docs/JOULESTUDIO/7e496fe6f81f4975aec07d9c1cbd2aef/01ffc9df21624c67995b365aa7880fc7.html",
     "note": "Known Issues says traces are kept for 90 days only; no setting to change this was found."
    },
    "evals": {
     "status": "documented",
     "url": "https://help.sap.com/docs/JOULESTUDIO/7e496fe6f81f4975aec07d9c1cbd2aef/7c688a84b2c74de8a132d1f641ee6178.html",
     "note": "Agent Evaluations covers running evaluation scenarios against a deployed agent, with metrics such as token counts and tool-call counts plus failure investigation; the page says evaluation runs are not compared across agent versions."
    }
   }
  },
  {
   "name": "Workday Agent System of Record (Workday agents)",
   "vendor": "Workday",
   "productUrl": "https://www.workday.com/en-us/ai-platform/agent-system-of-record.html",
   "scopeNote": "Assessed Workday Agent System of Record (Agent Management Hub, Agent Gateway) and Workday-built agents, including Sana for Workday, using the public Workday AI admin guide on doc.workday.com (PDF dated November 20, 2026); developer docs and Community content were not reviewed.",
   "scores": {
    "traces": {
     "status": "partial",
     "url": "https://doc.workday.com/admin-guide/en-us/workday-ai/agents/setup-considerations--agent-system-of-record.html",
     "note": "Agent input and output reports and an AI agent user audit trail are documented, but the input and output reports cover Workday-built agents only and no step-level trace of tool and model calls is described for customers."
    },
    "citations": {
     "status": "partial",
     "url": "https://doc.workday.com/content/dam/fmdita-outputs/pdfs/admin-guide/en-us/Admin-Guide-Workday-AI.pdf",
     "note": "Some agents document source attribution (for example, Payroll Agent can show which Workday data sources it queried, and the Deployment Agent gives links and citations), but this is agent-specific rather than a platform-wide feature."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Searched doc.workday.com and the Workday AI admin guide for data versions or retrieval timestamps recorded per agent run and found none."
    },
    "approvals": {
     "status": "partial",
     "url": "https://doc.workday.com/content/dam/fmdita-outputs/pdfs/admin-guide/en-us/Admin-Guide-Workday-AI.pdf",
     "note": "Sana for Workday asks users to review and confirm actions that need human-in-the-loop approval, and some agents submit changes for approval, but no general per-agent approval setting is documented in ASOR."
    },
    "scoping": {
     "status": "documented",
     "url": "https://doc.workday.com/admin-guide/en-us/workday-ai/agents/setup-considerations--agent-system-of-record.html",
     "note": "Admins set security groups per agent skill, each agent gets its own Agent System User account, tool execution also requires API security policy membership, and a View Security for Agent Skill report shows the required permissions."
    },
    "budgets": {
     "status": "not found",
     "url": null,
     "note": "The admin guide covers Flex Credit consumption tracking on the Consumption Usage page, but no per-agent or per-run spend, credit, or iteration caps were found."
    },
    "auditExport": {
     "status": "partial",
     "url": "https://doc.workday.com/admin-guide/en-us/workday-ai/agents/setup-considerations--agent-system-of-record.html",
     "note": "Agent audit trail reports and reportable data sources (such as Agent Input and Output Details) can be used in custom reports, but no SIEM or audit-log API export for agents is documented."
    },
    "retention": {
     "status": "partial",
     "url": "https://doc.workday.com/content/dam/fmdita-outputs/pdfs/admin-guide/en-us/Admin-Guide-Workday-AI.pdf",
     "note": "Fixed windows are documented (30 days for agent input and output details, 14 days for some agent chat histories), but no configurable retention for agent logs was found."
    },
    "evals": {
     "status": "not found",
     "url": null,
     "note": "Searched doc.workday.com and the Workday AI admin guide for built-in agent evaluation or test sets and found none."
    }
   }
  },
  {
   "name": "UiPath Agentic Automation (Agents, Agent Builder in Studio Web, Maestro)",
   "vendor": "UiPath",
   "productUrl": "https://docs.uipath.com/agents/automation-cloud/latest/user-guide/building-an-agent-in-studio-web",
   "scopeNote": "Assessed UiPath Agents built in Studio Web (the older standalone Agent Builder is being deprecated in favor of Studio Web), Maestro, and Automation Cloud governance features (AI Trust Layer, audit logs), using docs.uipath.com Automation Cloud latest.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://docs.uipath.com/agents/automation-cloud/latest/user-guide/agent-traces",
     "note": "Agent traces record each run as a timeline of steps, including LLM calls, inputs and outputs, errors, JSON payloads, token usage, and latency."
    },
    "citations": {
     "status": "documented",
     "url": "https://docs.uipath.com/agents/automation-cloud/latest/user-guide/best-practices-for-context-engineering",
     "note": "Agents with Context Grounding indexes can return citations, and the trace shows the retrieved chunks and citations; the Citations output supports only PDF and JSON data."
    },
    "lineage": {
     "status": "partial",
     "url": "https://docs.uipath.com/agents/automation-cloud/latest/user-guide/best-practices-for-context-engineering",
     "note": "Traces show which chunks were retrieved (with an outdated timestamp given as a sign of stale data), indexes show last-synced time, and UiPath recommends versioned index names, but no per-run record of the source data version is documented."
    },
    "approvals": {
     "status": "documented",
     "url": "https://docs.uipath.com/agents/automation-cloud/latest/user-guide/tool-guardrails",
     "note": "Custom tool guardrails can escalate to a human through an Action app before or after tool execution, and escalations pause the agent until a person resolves them in Action Center."
    },
    "scoping": {
     "status": "documented",
     "url": "https://docs.uipath.com/agents/automation-cloud/latest/user-guide/tool-guardrails",
     "note": "Each agent has an explicitly configured tool list, and guardrails can block a tool or filter its input and output fields; agents run under folder-assigned robot identities whose roles limit access."
    },
    "budgets": {
     "status": "documented",
     "url": "https://docs.uipath.com/automation-ops/automation-cloud/latest/user-guide/settings-for-studio-web-policies",
     "note": "Agents have Max. iterations (1 to 100, autonomous agents only) and max tokens per response settings, and Automation Ops policies can enforce maximum iterations, tokens, and temperature."
    },
    "auditExport": {
     "status": "documented",
     "url": "https://docs.uipath.com/automation-cloud/automation-cloud/latest/api-guide/download-audit-logs",
     "note": "Audit logs can be downloaded through organization- and tenant-level Audit Logs APIs, and Data Export can stream tenant audit logs to Azure Event Hub or AWS SQS; the AI Trust Layer audit tab has separate CSV export."
    },
    "retention": {
     "status": "documented",
     "url": "https://docs.uipath.com/agents/automation-cloud/latest/user-guide/agent-traces",
     "note": "Trace spans are kept 90 days by default, and admins can enable TTL in the AI Trust Layer policy to shorten retention to 1 or 7 days."
    },
    "evals": {
     "status": "documented",
     "url": "https://docs.uipath.com/agents/automation-cloud/latest/user-guide/agent-evaluations",
     "note": "Evaluation sets with LLM-as-a-judge, trajectory, and exact-match style evaluators are built in; runtime evaluations and Maestro evaluations are marked Preview."
    }
   }
  },
  {
   "name": "Palantir AIP (AIP Chatbot Studio, formerly AIP Agent Studio)",
   "vendor": "Palantir",
   "productUrl": "https://www.palantir.com/platforms/aip/",
   "scopeNote": "Assessed Palantir AIP on Foundry: AIP Chatbot Studio (renamed from AIP Agent Studio the week of April 27, 2026), AIP Logic, AIP Evals, AIP observability, and Foundry audit logging, using palantir.com/docs.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://www.palantir.com/docs/foundry/aip-observability/overview",
     "note": "AIP observability gives distributed traces across functions, actions, language model calls, and automations, plus logs with prompts and token usage, and chatbots have a View reasoning option."
    },
    "citations": {
     "status": "documented",
     "url": "https://www.palantir.com/docs/foundry/chatbot-studio/citations",
     "note": "Chatbots that use document or Ontology retrieval context get citations automatically, linking to Ontology objects, document pages, or external URLs; other context types need custom citation handling."
    },
    "lineage": {
     "status": "partial",
     "url": "https://www.palantir.com/docs/foundry/chatbot-studio/tools",
     "note": "Function tools can be pinned to a specific published version and observability ties into Workflow Lineage, but no per-run record of which data version an answer used was found."
    },
    "approvals": {
     "status": "documented",
     "url": "https://www.palantir.com/docs/foundry/chatbot-studio/tools",
     "note": "Action tools can be set to run automatically or only after user confirmation, and AIP Logic automations can stage Ontology edits as proposals for human review."
    },
    "scoping": {
     "status": "documented",
     "url": "https://www.palantir.com/docs/foundry/chatbot-studio/tools",
     "note": "Object query tools limit which object types and properties an agent can access, tools are configured per chatbot, and chatbots are filesystem resources with granular access control."
    },
    "budgets": {
     "status": "partial",
     "url": "https://www.palantir.com/docs/foundry/resource-management/llm-capacity-management",
     "note": "Admins can cap the share of tokens-per-minute and requests-per-minute each Project may use, per model, but no per-agent or per-run spend or iteration caps were found."
    },
    "auditExport": {
     "status": "documented",
     "url": "https://www.palantir.com/docs/foundry/security/monitor-audit-logs",
     "note": "audit.3 logs can be pulled into an external SIEM through audit API endpoints or exported to Foundry datasets."
    },
    "retention": {
     "status": "documented",
     "url": "https://www.palantir.com/docs/foundry/announcements/2025-05",
     "note": "Conversation retention is set per agent and version (24-hour default, or opt-in indefinite retention), and audit-log export datasets support a retention policy of up to 730 days."
    },
    "evals": {
     "status": "documented",
     "url": "https://www.palantir.com/docs/foundry/aip-evals/analyze-run-results",
     "note": "AIP Evals evaluation suites run test cases with LLM-as-a-judge or custom evaluators against Logic functions and published chatbots, with debug views and run comparison."
    }
   }
  },
  {
   "name": "Oracle AI Agent Studio for Fusion Applications",
   "vendor": "Oracle",
   "productUrl": "https://docs.oracle.com/en/cloud/saas/fusion-ai/aiaas/key-capabilities.html",
   "scopeNote": "Assessed Oracle AI Agent Studio for Fusion Applications (26D docs on docs.oracle.com); OCI Generative AI Agents was not assessed.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://docs.oracle.com/en/cloud/saas/fusion-ai/aiaas/monitor-agents.html",
     "note": "The Monitoring subtab lists every session with turns, status, and tokens, and a detailed trace view shows a step-by-step timeline with the tools called and per-step duration and metrics; metrics are gathered by a scheduled process, so they are not real time."
    },
    "citations": {
     "status": "partial",
     "url": "https://docs.oracle.com/en/cloud/saas/fusion-ai/26d/aiaas/how-do-i-use-ai-agent-studio.pdf",
     "note": "The guide mentions testing to validate sources cited by the agent and refers to document-tool citations in its FAQ, but the Add Document Tool page does not describe citation behavior."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Searched docs.oracle.com for per-run data versions or retrieval timestamps and found none; only draft and published versions of agents and workflows, and picking an agent version for evaluation, are documented."
    },
    "approvals": {
     "status": "documented",
     "url": "https://docs.oracle.com/en/cloud/saas/fusion-ai/26d/aiaas/how-do-i-use-ai-agent-studio.pdf",
     "note": "Tools have a Require Human Approval option so a person approves an action before the tool runs, and workflows support human approval nodes linked to published approval processes."
    },
    "scoping": {
     "status": "documented",
     "url": "https://docs.oracle.com/en/cloud/saas/fusion-ai/26d/aiaas/how-do-i-use-ai-agent-studio.pdf",
     "note": "Agents inherit Fusion role-based access controls, business object tools let admins choose which objects and fields to use or ignore, and Security tabs limit which roles can use workflows and plugins."
    },
    "budgets": {
     "status": "documented",
     "url": "https://docs.oracle.com/en/cloud/saas/fusion-ai/26d/aiaas/overview-of-ai-unit-budgeting-in-ai-agent-studio.html",
     "note": "AI unit budgets can be allocated to individual agents and agentic applications, with soft warning thresholds and an optional auto-stop that blocks new runs once the budget is used up."
    },
    "auditExport": {
     "status": "not found",
     "url": null,
     "note": "Searched docs.oracle.com for an audit log export or API covering AI Agent Studio agent activity and found none; only a budget change log and alert history in the Usage and Value tab were found."
    },
    "retention": {
     "status": "not found",
     "url": null,
     "note": "Searched docs.oracle.com for configurable retention of AI Agent Studio traces or conversation logs and found none."
    },
    "evals": {
     "status": "documented",
     "url": "https://docs.oracle.com/en/cloud/saas/fusion-ai/26d/aiaas/how-do-i-use-ai-agent-studio.pdf",
     "note": "Per-agent evaluation sets hold test questions, expected answers, and pass/fail metric thresholds, including an LLM-as-a-judge correctness score, and runs can be compared to catch regressions."
    }
   }
  },
  {
   "name": "Glean Agents",
   "vendor": "Glean",
   "productUrl": "https://docs.glean.com/agents/how-agents-work",
   "scopeNote": "Assessed Glean Agents (Agent Builder, headless builder, scheduled/independent agents) plus Glean admin console governance features documented on docs.glean.com.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://docs.glean.com/administration/agent-trace-export",
     "note": "Admins can export OpenTelemetry traces with spans for agent runs, tool executions and LLM calls (native tool arguments are redacted), and Agent Builder debug mode shows per-step inputs, outputs and timing."
    },
    "citations": {
     "status": "documented",
     "url": "https://docs.glean.com/user-guide/assistant/glean-chat/glean-chat-citations/glean-citations",
     "note": "Responses that retrieve company or web content show inline citation markers and a View sources list; answers from model knowledge alone carry no citations."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Agent/configuration versioning is documented, but no documentation was found recording the version or retrieval time of source data used in a run; only agent versioning is documented (up to 30 timestamped saved versions); no per-run record of source-document versions or retrieval times was found."
    },
    "approvals": {
     "status": "documented",
     "url": "https://docs.glean.com/tools/human-in-the-loop-experience-for-tools",
     "note": "Agents pause before write tools for user review and approval by default, and admins can exempt specific low-risk actions via a run-without-confirmation setting."
    },
    "scoping": {
     "status": "partial",
     "url": "https://docs.glean.com/administration/protect/ai-security/agent-access-policies",
     "note": "Agent access policies can block, filter or flag tool calls and outputs scoped to specific agents, users and tools, but the feature is labeled beta (agent identity with scoped credentials is also beta)."
    },
    "budgets": {
     "status": "documented",
     "url": "https://docs.glean.com/administration/management/usage/set-usage-limits-and-alerts",
     "note": "Admins can set a default or custom monthly credit limit per agent that can block further usage, and separate docs describe a hard per-run tool-call cap."
    },
    "auditExport": {
     "status": "documented",
     "url": "https://docs.glean.com/administration/management/audit-logs/admin-audit-logs",
     "note": "Admin audit logs export to CSV from the console, and Glean can configure ongoing delivery to customer destinations such as S3 or BigQuery on request (not self-serve)."
    },
    "retention": {
     "status": "documented",
     "url": "https://docs.glean.com/administration/assistant/configuration/chat-history",
     "note": "Admins set an org-wide chat history retention period (off, 30 days to 1 year) covering Assistant and agent runs; audit log retention defaults to 30 days and is extendable via Glean support."
    },
    "evals": {
     "status": "documented",
     "url": "https://docs.glean.com/agents/create-agents/create-agent-via-headless-builder",
     "note": "The headless builder supports creating reusable eval sets and running them against an agent with built-in task-completion and correctness checks, plus HTML/CSV result export."
    }
   }
  },
  {
   "name": "WRITER Agent and Writer Control Center",
   "vendor": "Writer",
   "productUrl": "https://dev.writer.com/home/introduction",
   "scopeNote": "Assessed WRITER Agent, playbooks and Agent Builder governed through Writer Control Center (the developer docs site is still titled Writer AI Studio; the 2025 launch branding was AI HQ).",
   "scores": {
    "traces": {
     "status": "partial",
     "url": "https://dev.writer.com/home/agent-observability",
     "note": "Event logs capture LLM calls with session, request and agent run IDs, but docs say they do not cover every platform action, tool-call logging is not described, and Agent Builder v2 events are excluded."
    },
    "citations": {
     "status": "documented",
     "url": "https://dev.writer.com/home/inline-citations",
     "note": "Knowledge Graph answers can include inline citations with a references object (optional in the API), and release notes say agent-mode deliverables include citations and a sources panel."
    },
    "lineage": {
     "status": "not found",
     "url": null,
     "note": "Searched dev.writer.com and support.writer.com for Knowledge Graph versioning or per-run source timestamps; nothing found beyond playbook version control."
    },
    "approvals": {
     "status": "documented",
     "url": "https://support.writer.com/articles/1496523599-get-started-with-playbooks",
     "note": "Playbook step approvals let builders assign an approver so the run pauses for sign-off before continuing (released September 2026)."
    },
    "scoping": {
     "status": "partial",
     "url": "https://support.writer.com/articles/7572520678-configuring-writer-agent-profiles",
     "note": "Admins can configure WRITER Agent profiles per team and enable specific connector tools for all users or selected teams, but controls are per team/connector rather than per agent."
    },
    "budgets": {
     "status": "not found",
     "url": null,
     "note": "Only spend notifications (alerts at a monthly dollar threshold, per October 2026 release notes) were found; no enforced per-agent or per-run limit was documented."
    },
    "auditExport": {
     "status": "documented",
     "url": "https://dev.writer.com/home/audit-events-export",
     "note": "An Audit Events Export API lets organizations poll admin and platform audit events by cursor for SIEM or compliance ingestion, with a 90-day default export window."
    },
    "retention": {
     "status": "documented",
     "url": "https://support.writer.com/articles/6397991324-setting-up-data-retention",
     "note": "Org admins on certain Enterprise plans can set automatic deletion (1, 3, 6 or 12 months) covering chat sessions including agent and playbook runs."
    },
    "evals": {
     "status": "partial",
     "url": "https://support.writer.com/articles/1496523599-get-started-with-playbooks",
     "note": "Playbooks support sandboxed test runs, synthetic-data testing and individual step testing, but no built-in evaluation sets or scoring were found."
    }
   }
  },
  {
   "name": "Sierra Agent OS",
   "vendor": "Sierra",
   "productUrl": "https://sierra.ai/product/agent-sdk",
   "scopeNote": "Assessed Sierra's Agent OS (Agent Studio and Agent SDK) using sierra.ai product pages and Sierra's own blog, since no public product documentation site was found and the Trust Center page returned no readable content.",
   "sourceNote": "No public product documentation site was found; scores rely on sierra.ai product pages and Sierra's own (first-party) blog posts.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://sierra.ai/blog/agent-traces",
     "note": "Sierra says every agent message generates a trace recording instructions, tool calls, knowledge lookups, network requests and per-step timing, in production and testing."
    },
    "citations": {
     "status": "not found",
     "url": null,
     "note": "Searched sierra.ai for customer-facing source citations in agent responses; only internal auditing of knowledge sources used was described."
    },
    "lineage": {
     "status": "partial",
     "url": "https://sierra.ai/blog/agent-development-life-cycle",
     "note": "Each agent release is described as an immutable snapshot including prompts, model versions and all available knowledge, but no documentation confirms the release is recorded per conversation."
    },
    "approvals": {
     "status": "partial",
     "url": "https://sierra.ai/blog/release-governance-guardrails-for-agents-at-scale",
     "note": "Organizations can require reviewer sign-off before agent changes are released, but approvals apply to changes rather than runtime agent actions (runtime escalation to humans is described separately)."
    },
    "scoping": {
     "status": "not found",
     "url": null,
     "note": "Searched sierra.ai for per-agent tool, data or action permission controls; only general mentions of role-based access and controlled system access were found."
    },
    "budgets": {
     "status": "not found",
     "url": null,
     "note": "Searched sierra.ai for per-agent or per-conversation usage, cost or turn limits; none found."
    },
    "auditExport": {
     "status": "partial",
     "url": "https://sierra.ai/blog/your-agent-laid-bare-and-why-it-matters",
     "note": "Conversation logs and performance data can be exported via OpenTelemetry, EventBridge, Pub/Sub or an export API, but no admin audit log export was documented."
    },
    "retention": {
     "status": "not found",
     "url": null,
     "note": "Searched sierra.ai and the DPA for configurable transcript or log retention; only contract-level deletion terms were found."
    },
    "evals": {
     "status": "documented",
     "url": "https://sierra.ai/blog/simulations-the-secret-behind-every-great-agent",
     "note": "Simulations auto-generate test cases and can be required to pass before releases, including in CI/CD pipelines."
    }
   }
  },
  {
   "name": "Zendesk AI agents",
   "vendor": "Zendesk",
   "productUrl": "https://support.zendesk.com/hc/en-us/articles/8966284087066-About-AI-agents-with-agentic-AI",
   "scopeNote": "Assessed Zendesk AI agents (customer-facing AI agents - Advanced, including agentic generative procedures), with auto assist and Zendesk Bridge noted where relevant.",
   "scores": {
    "traces": {
     "status": "documented",
     "url": "https://support.zendesk.com/hc/en-us/articles/8357749580186-Reviewing-conversation-logs-for-AI-agents",
     "note": "Conversation logs show per-message details including the agent's plan, the procedure step used and integration calls with request parameters and errors."
    },
    "citations": {
     "status": "documented",
     "url": "https://support.zendesk.com/hc/en-us/articles/9517744828058-Configuring-settings-for-generative-replies-in-AI-agents",
     "note": "Generative replies can display a Sources link listing the articles used (on by default), though CSV sources are not listed."
    },
    "lineage": {
     "status": "partial",
     "url": "https://support.zendesk.com/hc/en-us/articles/10040865503898-Managing-generative-procedures-for-AI-agents",
     "note": "Generative procedures keep timestamped version history and imported knowledge sources show last-updated times, but no per-conversation record of source versions was found."
    },
    "approvals": {
     "status": "partial",
     "url": "https://support.zendesk.com/hc/en-us/articles/9174548349978-About-actions-for-auto-assist-and-action-flows",
     "note": "Auto assist actions for human agents require agent approval unless admins pre-approve custom actions; approval controls for customer-facing AI agents were found only in the Zendesk Bridge early access program."
    },
    "scoping": {
     "status": "partial",
     "url": "https://support.zendesk.com/hc/en-us/articles/8357749301658-Connecting-knowledge-sources-to-power-generative-replies-in-AI-agents",
     "note": "Admins can connect or disconnect knowledge sources per AI agent and articles respect view permissions, but per-agent action/tool scoping is documented only for Zendesk Bridge (EAP)."
    },
    "budgets": {
     "status": "partial",
     "url": "https://support.zendesk.com/hc/en-us/articles/10479587390106-Managing-resolution-allowances",
     "note": "An account-wide Don't allow overage setting pauses AI agents when the resolution allowance is reached; no per-agent limit for customer-facing agents was found."
    },
    "auditExport": {
     "status": "partial",
     "url": "https://developer.zendesk.com/api-reference/ticketing/account-configuration/audit_logs/",
     "note": "The account Audit Logs API supports export (Enterprise plans), but docs do not confirm that AI agent configuration changes are included."
    },
    "retention": {
     "status": "documented",
     "url": "https://support.zendesk.com/hc/en-us/articles/8499219792154-Creating-deletion-schedules-for-bot-only-conversations",
     "note": "Admins can create deletion schedules for bot-only conversations and AI agent tickets in Admin Center."
    },
    "evals": {
     "status": "partial",
     "url": "https://support.zendesk.com/hc/en-us/articles/11126903515034-Using-the-conversation-simulator-to-test-your-AI-agents-EAP",
     "note": "A test widget supports manual testing and a conversation simulator runs simulated multi-turn scenarios, but the simulator is early access and no reusable test sets were found."
    }
   }
  }
 ]
}
