Archive preview
Historical essays show the thesis and workflow preview. The full examples, prompt pack, and downloadable packet unlock with a one-time purchase or archive pass.
The signal
June 11 discussion clustered around loop engineering: scheduled coding agents, worktrees, skills, subagents, durable state, and the shift from prompting an agent repeatedly to designing the system that prompts, checks, and escalates.
Why it matters
Most people still use AI agents like interns in a chat window. Loop engineering gives them a sane next step: turn one recurring task into a small operating system with a trigger, context, verifier, log, and human gate.
The useful shift is not that agents can run forever. That is the dangerous version. The useful shift is that some work is loop-shaped: it happens repeatedly, it has evidence to inspect, and the definition of done can be checked.
Workflow preview
- Pick one repetitive, reviewable, valuable task: PR babysitting, CI failure summaries, daily issue triage, changelog drafting, report QA, or recurring research scans.
- Write the loop contract: trigger, inputs, output format, allowed tools, blocked actions, budget, stop condition, and when to escalate.
- Give it durable context with an AGENTS.md, CLAUDE.md, SKILL.md, checklist, or rubric, plus a run log that outlives the chat.
Prompt preview
Design a report-only agent loop for this recurring task. Return the trigger, inputs, state file, allowed tools, blocked actions, stop condition, verifier rubric, budget cap, and human escalation rule. Do not assume the loop can write externally.
Where this goes wrong
- Do not build a loop for work that is not repetitive, reviewable, and valuable.
- Do not let the same agent grade its own work when the output matters.
- Do not run file-writing loops in your active checkout; isolate them first.
- Do not allow posting, pushing, merging, deleting, spending, or notifying without an explicit human gate.
- Do not ignore token and runtime caps. A stuck loop can become an open invoice.
Sources checked
- Addy Osmani's Loop Engineering essay framed the shift from prompting agents to designing the system that prompts themOpen source
- Firecrawl's June 11 explainer captured the practitioner debate and the repetitive/reviewable/valuable testOpen source
- OpenAI Codex Automations provide recurring background tasks, Triage, skills, and worktree isolationOpen source
- OpenAI Codex Skills package reusable workflow instructions in SKILL.md foldersOpen source
- Claude Code scheduled tasks document /loop and recurring prompt runs for PR babysitting, deploy polling, and long buildsOpen source
- Anthropic's dynamic workflows post describes fan-out, adversarial verification, and loop-until-done patternsOpen source
- The cobusgreyling/loop-engineering repo turns the pattern into checklists, starters, and production examplesOpen source
Loop contract template
- Loop readiness checklist
- PR babysitter prompt pack
- Maker/checker verifier rubric
- Run-log template with budget and escalation gates