agent-loop

You can automate a multi-agent loop by running the $loop command on a local document, which ingests sources and orchestrates a Codex-driven cycle of planning, execution, and verification. It processes notes and Markdown links to evolve code into executable plans.

The three-viewpoint research approach runs parallel lanes analyzing architecture dependencies, failure verification, and goal efficiency. These lanes synthesize their findings into a staged plan, ensuring bounded execution with triple challenge and verification before reassessing progress across cycles.

No, you cannot use this loop without explicit model and reasoning_effort parameters. The orchestration enforces a single strongest-model pin across all delegated lanes, requiring explicit runtime controls on every agent spawn during planning, execution, and run-decision transitions.

No, it is designed specifically for Codex runtimes that support delegated agent creation and explicit runtime controls. Standard runtimes lacking these explicit runtime control features will not be able to execute the bounded stages or manage the delegated multi-agent lanes.

The best way is to ingest flexible input forms like pasted backlogs and Markdown links into an automated software-improvement loop. The system researches the target, builds a staged plan, challenges it with three viewpoints, and executes in bounded stages with continuous refinement.

The single strongest-model pin is required to maintain consistency and discipline across all delegated execution lanes. By enforcing this pin alongside explicit reasoning_effort on every dispatch, the orchestration ensures coherent planning, execution, and run-decision transitions throughout the software-improvement cycle.