Caucus V1

Roles

The post calls Caucus the local web UI plus Python that calls Cursor's agent API and GitHub. That orchestrator is code in the runtime, not a third agent in this roster. The two agents are the long-lived implementation and review workers.

• Implementer

  Owns. Implements the change and opens or updates the pull request. On a later pass it updates the same pull request from review feedback.

  Uses the shared constraints above.

• Reviewer

  Owns. Reviews the pull request and either approves it or requests changes.

  Uses the shared constraints above.

Routines

• Minimal cycle

  On demand

  A control labeled along the lines of "Run minimal cycle" sends the task to the implementer, waits for structured output, sends context to the reviewer, and branches on approve versus request changes. The loop repeats until approval or a safety cap.

Guardrails

• The loop stops when the pull request is approved or a safety cap is hit. The UI also has a stop control.
• The run refuses to allocate fresh agents if the pair was never provisioned or a session died. It fails instead of swapping in a new worker.
• Each stage emits JSON the runtime can parse, including prUrl, reviewDecision, commentUrls, and actorClock. The orchestrator reads those fields.

Handoffs

The reviewer can send work back to the implementer. actorClock counts how many times each role has acted. When review requests changes, the orchestrator reads live GitHub comments with gh api and puts them in the next implementation task. The source of truth for that pass is the pull request thread, not the chat.

Results

What worked

The author reports a loop that completes from implement to review and back until the pull request is approved or the safety cap is hit. No metric is given.

What broke

To be added.

Lessons

To be added.

Prompts

To be added.

Source

• Christopher Meiklejohn, April 8, 2026

  https://christophermeiklejohn.com/ai/agents/reliability/zabriskie/2026/04/08/cursor-agents-caucus-v1.html

Notes