Problem
Currently, when a workflow is started, the agent harness and session information are not systematically managed during each step. There's no built-in mechanism for agents to check in regularly or for enforcing summary and completion validation at the end of steps/workflows.
Proposed solution
- On workflow start, retrieve agent harness and session info. Start a timer that checks back in on the session and sends a 'continue' message if the agent is stopped/inactive.
- When a step or workflow is completed, require a summary of what was done.
- Command should pass the criteria and summary with a prompt to validate whether the step/workflow is complete. If validation fails, send corrective info back to the agent describing what to fix/finish.
- Timer for workflow check-ins should be short, making it feasible as a background task for agents.
- Prefer solutions that avoid long-running background threads for the timer: explore leveraging start/complete/other workflow commands inside the agent harness loop. Seek a way to stream/return output back to the agent harness, ensuring the signal is strong and directs agent behavior appropriately.
- A simple mechanism could work (e.g., checking a file), as long as it's reliable and effective at steering agent(s) in the harness.
Alternatives considered
- Using plugins or external process/thread management (to be avoided if possible).
- Manual status checks without automated handling (not preferred).
Label: agent
Problem
Currently, when a workflow is started, the agent harness and session information are not systematically managed during each step. There's no built-in mechanism for agents to check in regularly or for enforcing summary and completion validation at the end of steps/workflows.
Proposed solution
Alternatives considered
Label: agent