On Agentic AI Coding Loops

A take, written from actually building the things — the fleet the Fleet Series describes. Opinionated by design.

The phrase is mostly vapor, but the thing underneath is real — and the fleet already is the real version, which is probably why the hype grates.

The loop itself is nothing. while not done: act; observe; decide. Twelve lines. Everyone demoing "agentic coding" has that part working. It is not where any of the difficulty or the value lives. What separates a toy from something you'd trust is everything wrapped around the loop — and it's exactly the stuff this harness obsesses over.

What actually matters

A real feedback signal. The best structural decision in the tooling here is generator–evaluator: one model produces, a different model grades against explicit criteria, and failures fold back into the next round. The book-research agent scoring five weighted dimensions against a threshold isn't "a loop" — it's a control system with an error term. Most agentic-loop demos have no grader at all; the model declares its own victory. That absence is the tell.

Gates, not autonomy. The hype word is "autonomous." The actual engineering is deciding what the loop may do unsupervised versus what queues for a human. Emitted tasks land as spec-needed and manual by default; the worker only touches ready + auto-approved + unblocked. That autonomy dial is the product. Anyone selling "just let it run overnight" hasn't drawn that line yet and will find out the hard way why it matters.

Convergence. A loop that re-runs has to not duplicate itself or thrash. Stable task identifiers so a re-derived finding dedup-skips, resumability so a long research run continues from the highest sprint — unsexy, and it's the whole difference between a reusable loop and a one-shot that demos great once on stage.

Verification is the game. Generation is cheap and getting cheaper by the month. Knowing whether the output is correct is the expensive, unsolved part — which is why the adversarial-verify / skeptic-panel patterns exist at all: a plausible-but-wrong result surviving is the real failure mode. Hype leads with generation ("it wrote 10k lines!"). Practitioners lead with verification. You can sort a room by which one someone reaches for first.

The limit nobody puts on the slide

Loops are great at bounded, verifiable work — a refactor with a test suite behind it, a mechanical migration, research-with-a-grader — and bad exactly where the spec is the hard part. The loop doesn't fix a vague spec, it amplifies one. That's why the book workflow's highest-leverage step is writing the questions, and why emitted tasks are review-then-implement rather than auto-merge. Encode the limit into the design instead of pretending it away.

The one question

When someone won't stop talking about agentic loops, this cuts through:

What's your evaluator, and where's the gate?

No crisp answer means they're describing a while-loop with vibes. A crisp answer means they're worth listening to.