Post n°004 — 2026.07.23 — 2 min
Loop Engineering
Coding agents work best inside a loop they can verify themselves. Designing that loop — what the agent starts with, what it can check, when it stops — beats prompting harder, every time.
The loop is the product#
When an agent underperforms, the reflex is to write a better prompt. The leverage is almost never there. It's in the loop the agent runs inside: what context it starts with, what it can verify without a human, and what makes it stop. Engineer those three and a mediocre prompt ships good work; skip them and the best prompt in the world ships confident guesses.
The inner loop#
Context in. A lane starts with the contract, the repo's instruction file, and the latest handoff — injected by a session-start hook, not pasted by hand. State lives in git, files, and handoffs, never only in a context window, so any lane can be killed and restarted fresh. Quality degrades measurably past ~100–150k tokens of context; the discipline that keeps you out of that zone is state-on-disk, not bigger windows.
Verifiers the agent runs itself. A commit gate in seconds, the eval subset, a verify skill that drives the real flows. Verification the agent can't run becomes your job at review time — so the cheapest missing verifier gets built as part of the task, not deferred. It feels slower and is faster by mid-session, because the agent self-corrects instead of queueing on your eyeballs.
Fan out for reads, single lane for writes. Parallel subagents are for investigation and review — heavy reading happens in someone else's window and only conclusions come back. Implementation runs single-lane with checkpoints unless workstreams touch disjoint modules and disjoint runtimes: worktrees isolate files, not simulators, ports, or datastores.
BENCHDJ Agenta generation loop with a critic and thresholds designed in from the startKnowing when to stop#
A lane ends at an evidence packet, a blocked-on-input question, or a spent retry budget — never at "this is probably right." And after two iterations on the same problem with no genuinely new hypothesis, the next move is fresh eyes: a critic with clean context reading only the repo and the spec. The session that produced the confusion is the worst place to resolve it.
The outer loop#
The same discipline runs on a longer clock. Every morning at 06:17, every focus repo's gate and eval suite runs while I sleep; a digest lands, and the session-start hook serves it with the first coffee. Weekly, a proposer agent mines the digests and logs for recurring friction and files improvement proposals. On a model release: requalify by rerunning the suites and comparing — never switch on launch-day vibes. Drift gets caught by machinery, on a schedule, not by someone noticing.
BENCHJIMnightly baseline-compare: a regression is a verdict, not a feeling