Six concrete agent loops for coding workflows, with public configs
Cameron Westland shares six production-ready "loops that prompt coding agents" — scheduled and event-driven automations — along with public configs and skills in response to viral calls to move beyond manual prompting.
Score breakdown
The post provides the first concrete, public implementation of the "design loops, not prompts" pattern that Steinberger and Cherny described but never demonstrated, giving practitioners actual configs and skills to study or reuse.
- 01Peter Steinberger's tweet urging developers to design loops that prompt agents reached eight million views.
- 02Boris Cherny, creator of Claude Code, said on Acquired Unplugged: 'I don't prompt Claude anymore. I have loops running.'
- 03Westland has been running loops for months, primarily as Codex app automations, with configs published in `camwest/agent-skills`.
Cameron Westland's post responds directly to two viral moments: Peter Steinberger's tweet urging developers to design loops that prompt agents (which reached eight million views), and Boris Cherny's statement on Acquired Unplugged that he no longer prompts Claude himself — his loops do. Neither showed an actual loop, so Westland provides six concrete ones, backed by public configs and skills in the `camwest/agent-skills` repo.
Westland's own loops are primarily the second kind, with experiment-style inner loops triggered once they wake up.
The post first clarifies that "loop" means at least three different things: an autonomous task loop ("keep going until done"), exemplified by Geoffrey Huntley's Ralph pattern (`while :; do cat PROMPT.md | claude-code ; done`) and the `/goal` command in Codex and Claude Code; a scheduled or event-driven loop that runs without a human present, exemplified by `HEARTBEAT.md` in OpenClaw (a checklist reconsidered every 30 minutes) and its descendants in Codex automations and Claude Code routines; and orchestration fan-out, described as MapReduce with agents closer to the actor model. Westland's own loops are primarily the second kind, with experiment-style inner loops triggered once they wake up.
The "gateway drug" loop is `graphite-pr-babysit`: the agent watches a PR's review run and CI, fixes what it can, and returns only when the PR is clean or stuck — a pattern Westland describes as generalizing to any workflow that watches for a state change in an external system. A second class of loops runs inner experiment cycles, such as a previously described autoresearch loop that ran 49 experiments in an hour against a slow Python path, reducing p95 latency from 339ms to 34ms at a cost of $24. The same shape is now applied to testing agent behavior in production, using a skill called `investos-runtime-test-loop`. The source text is truncated before all six loops are fully described.
Key facts
- 01Peter Steinberger's tweet urging developers to design loops that prompt agents reached eight million views.
- 02Boris Cherny, creator of Claude Code, said on Acquired Unplugged: 'I don't prompt Claude anymore. I have loops running.'
- 03Westland has been running loops for months, primarily as Codex app automations, with configs published in `camwest/agent-skills`.
- 04The post distinguishes three types of 'loop': autonomous task loops, scheduled/event-driven loops, and orchestration fan-out.
- 05Geoffrey Huntley's Ralph pattern is cited as the purest autonomous loop: `while :; do cat PROMPT.md | claude-code ; done`.
- 06The `graphite-pr-babysit` skill watches a PR's review and CI, fixes what's fixable, and escalates when stuck.
- 07A prior autoresearch inner loop ran 49 experiments in an hour, cutting p95 latency from 339ms to 34ms at a cost of $24.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 11, 2026 · 08:34 UTC. How this works →