Loop-Machine

"Most agent systems die not from a weak model but from a weak harness."
— Andrej Karpathy, LOOPS.md

Loop-Machine replaces one-shot agent delegation with a three-role, contract-negotiated, crash-resumable loop. For any non-trivial task, spawn three agents instead of one: a Planner, a Generator, and an Evaluator. The orchestrator monitors progress and only intervenes when the contract itself is wrong.

Why Three Agents?

When a single agent does everything — plan, execute, and evaluate — it becomes sycophantic. It rubber-stamps its own work, converges on slop, and never truly verifies. Separating the roles forces genuine quality at every gate.

The Three Roles

Protocol

1. Spawn Planner. Give it the task. It writes contract.md with ~27 testable assertions, constraints, and a tool plan. It does NOT execute.
2. Approve contract. Read contract.md. If the criteria are vague or too few, push back. Once approved, the contract is frozen.
3. Spawn Generator. Give it the contract path. It reads the criteria, produces the deliverable, writes progress.md and log.md. It does NOT evaluate itself.
4. Spawn Evaluator. Told from the start: "The output is almost certainly broken. Prove it." Reads the contract and the Generator's output. Writes evaluation.md with per-criteria pass/fail, scores, and gap explanations.
5. Monitor & Decide. Read evaluation.md. Three outcomes:
   • All PASS → Task complete. Deliver summary.
   • Some FAIL, retries remain → Delete Generator output. Go to step 3. Let the loop restart.
   • Contract is wrong → Escalate. The contract needs revision, not the build.

File Structure

delegations/[task-slug]/
  contract.md        # Planner's spec + testable criteria (~27 assertions)
  progress.md        # Generator's current state
  log.md             # Append-only iteration history
  evaluation.md      # Evaluator's pass/fail per criterion
  output/            # Generator's deliverable files
  trace/             # Raw session transcripts (for debugging)

The system must be able to crash, lose its session, and resume by reading only contract.md, progress.md, and log.md.

Contract Template

# Contract: [Task Slug]

## Objective
[One-sentence task description]

## Testable Criteria (aim for ~27)
- [ ] Criterion 1: [exact, falsifiable statement]
- [ ] Criterion 2: [exact, falsifiable statement]
...

## Constraints
- Must not: [forbidden approaches]
- Budget: [max iterations or time]
- Output format: [exact format spec]

## Subjective Rubric (if applicable)
| Axis    | Weight | 3 Good Examples | 3 Bad Examples |

## Exit Conditions
- All criteria PASS → SUCCESS
- [3] consecutive restarts → ESCALATE
- [N] total iterations → REPORT partial

Retry Budget

Default: 3 retries per task. After budget exhausted with persistent failures → escalate. Counter resets if a retry produces an all-PASS evaluation. Track in log.md.

Origin

This skill is based on Andrej Karpathy's LOOPS.md: Field Notes on Agents That Run for Days, which distills nine principles for building autonomous agent loops. The key insight: "Write the loop, not the prompt" — the unit of leverage is the procedure, not the message.