Autonomous loops

Autonomous Loops

A loop is an agent that prompts itself — selecting the next task, doing it, checking it, logging progress, repeating — until a stop condition is met. Done well it multiplies throughput; done badly it manufactures comprehension debt (code nobody reviewed) and invites cognitive surrender (accepting output without thinking). The job of this skill is to build loops that stay accountable: a verifiable stop, a checker separate from the maker, a place for decisions that need a human, and hard caps on spend — with you still the engineer who confirms it works, not the person who pressed go.

This is the autonomy conductor. It composes [[parallel-subagents]] (fan-out), [[long-running-agents]] (multi-hour coherence), and [[agent-verification]] (trustworthy "done"); it does not replace them. It sits under [[agent-guardrails]] — autonomy never widens what an agent may do without approval.

When to Use

• Setting up an agent to run unattended or on a schedule (discovery, triage, recurring fixes)
• Building a self-prompting cycle (/goal until a stop condition; /loop on a cadence)
• A backlog of similar, well-specified tasks an agent could grind through
• You want automation to surface decisions, not make irreversible ones

Skip / don't loop when:

• The task is one-off or ambiguous — direct prompting is still the right tool; don't loop for its own sake
• Success can't be verified programmatically — without a checkable stop, a loop just drifts confidently
• The work is high-blast-radius (prod data, money, security, irreversible) — keep a human in the turn
• You'd be unable to review the output volume the loop produces — that's comprehension debt by design

Not a substitute for your judgment: a loop is a tool built with critical thinking, not a way to avoid it.

Process

1. Write a verifiable stop condition first

Before any loop runs, define done as something a machine can check — tests green, a metric met, an empty queue, a checklist all-passed ([[agent-verification]] owns the check). "Until it looks done" is not a stop condition; it's an open tab burning tokens. No verifiable stop → don't loop.

2. Split maker from checker

The agent that does the work must not be the only one that judges it — models grade themselves too generously. Route each iteration's output through a separate evaluator (a sub-agent, ideally a different model) per [[agent-verification]]. The checker can refuse and send work back.

3. Bound the blast radius

• Run in isolation — git worktree / branch, sandbox — and integrate via pull request, never direct pushes to a protected branch ([[git-flow]]).
• Keep [[agent-guardrails]] active: destructive/secret/prod actions still require explicit approval, loop or not.

4. Add circuit breakers

Set explicit caps before starting: max iterations, wall-clock limit, and a token/cost ceiling ([[finops-budget]] for spend). On breach, stop and surface — a runaway loop is a budget incident.

5. Route human decisions to a triage inbox

Anything the loop shouldn't decide — ambiguous trade-offs, risky changes, judgment calls — goes to a single triage inbox (an issue list, a Linear board, a TRIAGE.md) for a human, instead of the agent guessing. Confident guessing on missing intent is intent debt; make the loop ask, not assume.

6. Make progress durable and resumable

Each iteration commits a meaningful unit and updates an external progress/plan file so a reset or crash can resume without redoing work ([[long-running-agents]] owns the mechanics).

7. Supervise — stay the verifier

Sample the loop's output continuously; don't wait for the end. Your job is to ship code you confirmed works. If you find yourself rubber-stamping, that's cognitive surrender — slow the loop, shrink the batch, or stop it.

8. Know when to stop the loop

Stop when the stop condition is met, a circuit breaker trips, the triage inbox is backing up, or the checker's reject rate climbs (the loop is thrashing). Then review the batch as a human before merge.

Common Rationalizations

• "It's autonomous, I don't need to review it." — Then you're shipping unreviewed code; that's comprehension debt, and it's yours when it breaks.
• "The agent said it's done." — Self-grading is biased; a separate checker decides done ([[agent-verification]]).
• "I'll set a budget later." — Later is after the runaway bill; caps go in before the first iteration.
• "Loop everything to save time." — Direct prompting still wins for one-off or fuzzy work; looping those wastes tokens and trust.
• "The agent can decide that edge case." — If it needs judgment, it goes to triage; confident guesses become intent debt.
• "I'll just press go and check at the end." — Sampling continuously catches drift while it's cheap.

Red Flags

• A loop with no machine-checkable stop condition
• The maker and the checker are the same agent (or same model grading itself)
• No iteration / time / token ceiling set
• Direct commits to a protected branch instead of PRs from isolation
• No triage inbox — the agent guesses on ambiguity instead of asking
• Output volume exceeds what any human will actually review
• Reject rate climbing while the loop keeps running (thrashing)

Verification

• A verifiable, machine-checkable stop condition was written before the loop ran
• Maker and checker are separate (ideally different models); the checker can reject ([[agent-verification]])
• Loop runs in isolation and integrates via PR; [[agent-guardrails]] still gate risky actions
• Iteration, wall-clock, and token/cost circuit breakers set ([[finops-budget]])
• A triage inbox exists for decisions the loop must not make
• Progress is committed and resumable each iteration ([[long-running-agents]])
• A human sampled output during the run and reviewed the batch before merge — not rubber-stamped