Skip to content

Execution Contracts And Result Capture

The runtime does not treat “the model stopped responding” as success.

That is one of the main differences between Agirunner and a simple “LLM plus shell” loop.

When the runtime claims a task, it receives an explicit contract from the platform that shapes execution:

  • loop mode
  • model and provider context
  • tool permissions
  • budgets and limits
  • completion and reporting expectations

The runtime enforces that contract instead of inferring workflow meaning from role names, stage labels, or freeform model prose.

During and after execution, the runtime captures:

  • structured outputs
  • artifacts
  • logs
  • metrics
  • execution metadata

Those outputs are then shipped back to the platform so operators can inspect evidence instead of inferring success from a raw transcript.

Contract-aware completion is what lets Agirunner tell the difference between “the model stopped” and “the task satisfied the system’s actual output expectations.”