loop-engineer

To prevent false completion in agent loops, route broad objectives into specific design, contract, run, repair, eval, or flywheel phases with explicit success criteria, verification gates, and terminal states. This ensures outcomes are verified before being treated as done.

Designing a repair loop involves routing the failing run to a dedicated repair phase that applies durable on-disk artifacts and verification gates. This setup allows the loop to survive context loss, resume execution, and be audited later without restarting from scratch.

To audit an inherited agent loop, route the inspection request to the correct workflow phase to check for durable on-disk artifacts, explicit success criteria, and terminal states. This identifies missing verification gates and prevents false completion in the existing system.

Yes, setting up runtime monitoring for agent systems requires explicit success criteria, verification gates, and terminal states. These durable elements ensure the loop can be properly audited, survive context loss, and avoid falsely marking an unverified outcome as complete.

Agent loops falsely report completion when they treat unverified outcomes as done. By routing objectives through distinct contract, run, and eval phases with verification gates and terminal states, you enforce durable checks that prevent unverified results from closing the loop.