Unattended loops don’t have a token problem — they have a feedback problem

Signadot’s Sep 16 essay is the cost-economics half of always-on factories. “Write loops, not prompts” means a small program hands the agent a goal, runs checks, feeds failures back, and retries until pass or budget — unattended, long-running, often parallel. That is exactly the shape of a spend problem. Interactive sessions had a human who could close the laptop. A loop removes that person by design; whatever one hour of watched agent spend looked like, a dozen overnight loops can burn without anyone noticing.

Two identical loops against the same repo can still produce very different bills. The separator is not the loop code — it is what the loop can observe. Guardrails (iteration caps, no-progress detection, spend ceilings) are necessary and insufficient: they bound how much a wasteful loop can waste; they do not make it less wasteful. Total loop cost is a product, not a sum: iterations needed × cost per iteration. Halving both quarters the bill. A floor under either term is a floor under the total no cap can lift.

Feedback quality sets the iteration count. A bare fail forces guessing; structured error from the real system localizes and moves on. Where the loop closes sets price per iteration: CI after the change leaves the session buys minutes-to-hours of pipeline queue and leaves the agent iterating on partial signals in between; a runtime the agent can call inside the session lands in seconds. Pulling high-fidelity feedback into the inner loop is the only move that shrinks both terms at once.

Distributed systems push both numbers the wrong way. Local tests and mocks are fast and low fidelity; staging/CI are high fidelity and slow. Loops pay for both, so neither end of the tradeoff is good enough at agent concurrency. The platform-shaped answer is a verification surface the organization owns: a runtime reachable in seconds (ephemeral environments that deploy only touched services and route request-keyed traffic through a shared baseline), a declarative definition of done agents do not invent, feedback detailed enough to localize faults, and controls that make spend forecastable. Application developers state checkable goals; platform engineers own what every loop passes through — the same CI/CD transition, with the new requirement that it works during authoring, before a PR exists, at whatever concurrency the loops generate.

Why it matters for code review

Harnessed overnight momentum is not “more tokens while you sleep.” It is a verification surface cheap and faithful enough that continuous loops stop thrashing. Review bandwidth under sustained load is better spent on diffs that already failed closed against production-like evidence — not on reconstructing what a self-scoring overnight agent believed was done.