# ICM: Loop Engineering (knowledge)

Reference knowledge on "loop engineering" - the mid-2026 way of building autonomous AI agents - and how The ZAO applies it. Full analysis: ZAOOS research doc 994. Core idea: you don't prompt agents, you design the loops that run them.

## The 5 loops (name which one you mean before arguing about autonomy)
- Execution loop - one task: call tool, read result, decide, repeat; ends on env feedback (tests, API).
- Task loop (the "Ralph loop", Geoffrey Huntley) - restart a fresh-context agent on the SAME spec over and over; re-feeding the spec each time prevents context rot. Ends on spec + tests.
- Product loop (the "software factory": Factory, Warp Oz, Anthropic internal) - iterates a whole codebase + backlog: triage, spec, implement, review, verify, ship, monitor. Signals come from outside (issues, logs, feedback).
- System loop (autoresearch: Karpathy, Meta) - the outer loop improves the system itself: prompts, harnesses, model, and the evals. "The loop is the product."
- Oversight loop - set goals, allocate budget, cull work. Exit condition = a human. "Inner loop is capability; outer loop is agency."

## Core principles
- Autonomy is a separate dial on EVERY loop, not on/off. The real question is "what signal sets this dial," not "auto or not."
- A loop without a wired-in stop signal does not converge - it runs until money, rate limit, or a human stops it. Every loop needs a nameable exit + a hard cap.
- Fan-out (dispatch -> gather -> validate) is a PIPELINE, not a loop - no feedback = "just a for statement."
- The bottleneck is human REVIEW, not agent capability (even Anthropic is review-bound).

## The Karpathy method
- A loop = a goal the AI works toward until met: discover, plan, do, check, feed back, repeat. Three parts: clear goal, a feedback signal, a hard stop (goal met OR after N tries). Origin: Karpathy's AutoResearch (Mar 2026).
- When a loop is a mistake: it earns its cost only when the task is real, repeatable, checkable, and worth the tokens. On limited-token plans a heavy loop hits the wallet before the payoff.

## How The ZAO maps
- Claude Code session = execution loop. ZOE fix-PR pipeline = task/Ralph loop. ZOE orchestrator = product loop. /loop = system loop. Zaal on the board = oversight loop.

## Related boxes
- ZAO Assistant (zao-assistant)
- The ZAO (thezao)

## Deeper (primary sources, DEEP pass)
- Karpathy AutoResearch ("the Karpathy Loop"): propose -> execute -> evaluate -> commit/rollback; the model cannot override the stop; git is ground truth.
- Ralph loop (Huntley): fresh agent per iteration vs long sessions - higher token cost, lower error accumulation.
- Inner vs outer (Osmani): inner = capability, outer = agency; most teams only invest in the inner.
- LangChain four-loop: agent work -> verification -> event/analysis -> improvement; value compounds in the outer two.
- Software factories (Warp Oz, Factory): ratchet auto-merge as trust grows (e.g. 20% -> 60%).
- Critique camp: understanding is the real bottleneck (Litt); "there is no auto" - agents do 80% grunt, humans the 20% (Bakaus); step DOWN an abstraction, prefer deterministic control loops (Horthy).
- Failure mode to avoid: "loopmaxxing" - adding loops to vague goals with no eval/stop.

## Loop checklist (build a good loop)
- Pre: clear goal, a rubric/eval for done, context strategy (fresh vs compacted), a token/cost budget.
- During: an exit signal the model cannot fake (tests/eval/git), an iteration cap, a cost ceiling, no-progress detection.
- Post: a human checkpoint sized to risk, an audit of changes, and a review-throughput plan (review is the real bottleneck).

Full analysis: ZAOOS research doc 994 (DEEP).