Bounded Execution

AgentRail runs your agent through a bounded loop, not an open-ended chat. Each issue moves through three phases — plan → execute → verify — and the run is scoped to one human-defined issue. The agent decides how to complete it; it never invents its own goals.

The three phases

• Plan — the agent reads its compiled context pack and lays out the change it intends to make.
• Execute — it makes the smallest change that satisfies the issue's acceptance criteria.
• Verify — the work is checked against the Objective Gate: tests, build, and lint pass, and the acceptance criteria are met. That gate — not the agent's opinion — is the only signal that says the run is done.

Automatic retry on failure

When verification comes back red, the loop doesn't stop dead and it doesn't retry forever. It re-attempts with a compacted handoff of what failed. If it still can't reach a passing gate within the issue's budget, it stops and routes to a human rather than burning tokens indefinitely. See Review & Gates for the terminal states and AFK for the per-issue budget leash.

Running an issue

agentrail run issue 42 --agent claude

This runs the full plan → execute → verify loop for issue 42, retrying on verification failure, and stops at review. Use --agent codex to run Codex instead, or agentrail run to run the workflow without a specific issue number.