loop-engineer

A Claude Code skill + subagent for setting up and administering agentic loops — the 2026 "loop engineering" paradigm where the developer's object of attention shifts source code → agent → loop.

"I don't prompt Claude anymore. I have loops running that prompt Claude… My job is to write loops." — Boris Cherny, creator of Claude Code (Anthropic)

The loop primitives now exist in Claude Code (claude -p, /loop, /goal, Tasks, cloud Routines), but nothing helps you choose, scaffold, and guard a loop. This does. Every loop it produces ships with non-negotiable guardrails so it can't "loopmaxx" — run forever against a vague objective and burn money.

What's inside

.claude/
├── skills/loop-engineer/        # the knowledge pack
│   ├── SKILL.md                 # decision tree + 5-point guardrail checklist
│   ├── reference/
│   │   ├── primitives.md        # claude -p, /loop, /goal, Tasks, Routines, cost gating
│   │   ├── taxonomy.md          # the 5 patterns + when to use each + nesting
│   │   └── guardrails.md        # failure modes + mitigations (with the real numbers)
│   └── templates/               # runnable, guardrailed scaffolds
│       ├── headless-loop.sh         # single agentic loop around `claude -p`
│       ├── evaluator-optimizer.sh   # generator + SEPARATE critic
│       ├── meta-prompt-refine.sh    # a loop that refines the prompt another loop runs
│       ├── fanout-orchestrator.sh   # orchestrator-workers fan-out + verify
│       └── goal-loop.md             # first-party /goal and /loop recipes
├── agents/
│   ├── loop-architect.md        # interactive loop designer (scaffolds)
│   └── loop-teacher.md          # interactive tutor (explains how to think in loops)
└── commands/                    # slash-command shortcuts
    ├── new-loop.md              # /new-loop <goal>   → design + scaffold a loop
    ├── loop-audit.md            # /loop-audit <file> → audit a loop vs. guardrails
    └── learn-loops.md           # /learn-loops       → learn loop engineering interactively

Note: Claude Code already ships a built-in /loop command (interval/self-paced repetition), so these shortcuts are namespaced /new-loop and /loop-audit to avoid the collision.

The five patterns

Nest them: an orchestrator can spawn evaluator-optimizer inner loops; a meta-loop wraps a headless loop and rewrites its prompt between runs.

The one rule that prevents most disasters

No binary, verifiable exit condition → no loop. "Improve the UX" has no pass/fail and produces infinite loops + large API bills. "Make npm test exit 0" does. A real 4-agent loop once ran 11 days and cost ~$47,000 for lack of this.

Every scaffold enforces the five-point guardrail checklist:

1. Verifiable exit condition — checked each iteration.
2. Max-iteration cap — a hard bound in code, not in the prompt.
3. Budget cap in code — sum total_cost_usd, stop before the next call. (Alerts are not enforcement.)
4. Sandbox — worktree/container/branch, never unattended on main.
5. Human checkpoint — for any push/deploy/send/irreversible action.

Install

Drop the skill and agent into a project's .claude/, or into ~/.claude/ to make them available everywhere:

# project-local
git clone https://github.com/fltman/loop-engineer.git
cp -r loop-engineer/.claude/skills/loop-engineer  /path/to/your/project/.claude/skills/
cp -r loop-engineer/.claude/agents/loop-architect.md /path/to/your/project/.claude/agents/

# or globally
cp -r loop-engineer/.claude/skills/loop-engineer  ~/.claude/skills/
cp -r loop-engineer/.claude/agents/loop-architect.md ~/.claude/agents/

Use

• Just describe the loop and the loop-engineer skill triggers and scaffolds it: "set up a loop that runs until the tests pass, max 10 iterations, under $5."
• Design it interactively with the loop-architect subagent — it interrogates the goal, forces it to be binary, picks the pattern, scaffolds the files, and wires the guardrails.
• Learn the mental model with the loop-teacher subagent — it teaches loop engineering Socratically against your own task (no scaffolding, just understanding), then hands off to loop-architect when you're ready.
• Use the slash shortcuts for comfort:
  • /new-loop <goal> — design + scaffold a guardrailed loop for a goal
  • /loop-audit <file> — audit an existing loop against the five guardrails
  • /learn-loops — learn how to think about loops interactively

Run a scaffold directly, e.g.:

EXIT_CHECK='npm test --silent' \
TASK_PROMPT='Make all unit tests pass. Run the suite to verify.' \
MAX_ITERS=10 BUDGET_USD=5.00 \
.claude/skills/loop-engineer/templates/headless-loop.sh

The templates parse claude -p --output-format json. Field names (result/response, per-call vs. cumulative total_cost_usd) can vary by version — verify against your Claude Code before unattended runs.

Provenance & honest attribution

Distilled from a multi-source, adversarially fact-checked research pass. The architecture taxonomy comes from Anthropic's Building Effective Agents; the meta-loop lineage from Self-Taught Evaluators (Wang et al., arXiv 2408.02666). Shell fan-out / headless-CI loops are community patterns, not official Anthropic endorsements — and are labeled as such throughout.

License

MIT — see LICENSE.