EXPLAINER / EXECUTION AGENTS
How to set up
an execution agent.
A practical guide to turning an analytical assignment into data access, tools, budget limits and an output format, using a regional employment comparison as the example.
THE EXAMPLE REQUEST
Start with a specific analytical request.
“Compare employment across these regions and prepare a table with sources.” That request needs a definition of employment, a geographic level and a period before the agent can choose suitable data.
The execution-agent configuration sets out the assignment, the tools it can use and the permitted output formats. The platform’s execution contract requires both a tool it can run and an output.
REQUEST TO TOOL GRANT
Match the request to the tools.
| Requirement | Permitted operation | Result needed |
|---|---|---|
| Find the measure | Discover measures in an approved source scope. | The selected metric and its coverage. |
| Compare the region set | Compare geographies for that metric and period. | The returned observations for each eligible geography. |
| Explain the difference | Calculate an appropriate change, ratio or ordering over comparable results. | The calculation and its evidence references. |
| Return the work | Write the permitted report or structured JSON. | The table, interpretation and unresolved inputs. |
LIMITS ARE PART OF THE DEFINITION
Set access and budget limits for the assignment.
The agent manifest records sources, allowed actions, tool limits and budget settings. Workflow permissions determine which roles the agent can take in a process.
Adding a source or an external action changes what the agent can access or do. Review those permissions when the assignment changes.
OUTPUT CONTRACT
Choose the output for the person or system using it.
A reader needs the findings and their basis. An application may need typed rows, reference identifiers and a completion status. Agreeing the format at setup lets the agent’s results fit the surrounding software.
See the execution-agent capability pageAPPLICATION CONTEXT
Use this component within a wider application.
An agent’s structured output can become an input to an application screen or a later workflow step. The custom build supplies the account mapping, domain logic and user actions around that result, while Evidence OS provides the selected analytical services.
See the platform extension model