agent-loop

agent-loop is a lightweight CLI for designing and running bounded AI coding agent loops. The first MVP focuses on a single high-value workflow: fix-ci.

The tool treats a loop as five explicit parts:

• trigger: what starts the loop.
• context: what the agent should inspect.
• agent: the coding agent adapter.
• verifier: the command that proves progress.
• state: durable journals and reports for every run.

Install

cargo install --path .

Usage

Create the built-in loop spec:

agent-loop loop init --verifier "cargo test"

Prepare a fix-ci agent prompt from a failing verifier without invoking the agent:

agent-loop loop run fix-ci --verifier "cargo test" --dry-run

Run fix-ci for real:

agent-loop loop run fix-ci --verifier "cargo test"

Real runs create an isolated git worktree under .agent-loop/loops/... before invoking the agent. Dry runs write the same journal/report artifacts but do not create a worktree or call the agent.

Inspect local loop runs:

agent-loop loop status
agent-loop loop inspect <run-id>

Promote a useful run into a reusable skill note:

agent-loop loop promote <run-id> --as-skill fix-rust-ci

Current Scope

• Built-in fix-ci loop template.
• Claude Code adapter using claude -p.
• Shell verifiers.
• Hard iteration, time, cost, and no-progress budget fields in loop specs.
• JSONL run journal and markdown report for every run.
• Git worktree isolation for non-dry-run execution.

Development

cargo test