loop-engineering

To stop agent loops from endless retries, you design bounded agent workflows with explicit triggers, finite iteration caps, time budgets, and independent verification before execution. This approach replaces self-grading with evidence-driven contract validation to prevent runaway automation.

Designing a CI repair automation with stop conditions involves converting the task into a bounded maker-checker loop. You define the trigger, objective, scope, evidence, and finite budget, then use script-based validation to ensure the repair halts when limits are reached or recovery succeeds.

A bounded agent loop is a controlled, repeatable workflow with finite iteration and time caps. You need one for multi-step tasks like scheduled automation or bounded research, where tasks require explicit state, permission boundaries, and independent verification to prevent drift or unsupported state mutations.

For tasks requiring independent verification, a single-agent topology is generally insufficient. You should select a maker-checker or manager-workers pattern to separate execution from validation, ensuring the loop contract is independently verified and preventing self-verified claims.

Limitations of bounded loops include detecting and handling stalled loops, vague stop conditions, and unsupported writes. The workflow requires a durable contract and script-based validation to prevent these failures from wasting time or mutating shared state during scheduled execution.