In early June, Boris Cherny — the head of Claude Code at Anthropic — said something that stuck: "I don't prompt Claude anymore. I have loops running that prompt Claude… My job is to write loops." A few days later Peter Steinberger, creator of OpenClaw, compressed it into twelve words: stop prompting your coding agents, start designing the loops that prompt them. And the next day Addy Osmani gave the idea a name and an anatomy: loop engineering.
If you build agents for a living, the idea itself wasn't new — agents have been while-loops around a model since ReAct, and half of what's now called a loop, an SRE would call a reconciliation controller with better manners. What changed in June is that the industry agreed on where the leverage lives. Not in the prompt. Not even in the context. In the loop: the triggers, the verification, the budgets, and the stop rules wrapped around the agent. A prompt can't express "keep going until the tests pass, spend at most this much, and stop and ask if you touch billing." A loop can.
This is the third layer of a story we've written about before — context, then the harness, now the loop that drives the harness. And I want to write about it honestly, because most of what's been published since June is either euphoria or a eulogy, and the truth is more useful than both.
The skeptics have a point — three, actually
The backlash arrived within three weeks, and it was not stupid. The Register's headline said it plainly: loop engineering still needs humans in the loop. The 1,800-comment Hacker News thread said it more rudely. The serious objections reduce to three:
Unattended loops make unattended mistakes. Even Osmani, who named the practice, warns exactly this. A loop that runs while you sleep also fails while you sleep — and verification, in the end, remains a human responsibility.
Cost. Agent workloads are reported to consume several times the tokens of ordinary chat use, and multi-agent setups multiples of that. Bloomberg reported that Uber capped every employee at $1,500 a month in AI spend after burning through its annual budget in about four months. Critics have pointed out, fairly, that the loudest loop advocates work at companies that sell the tokens the loops consume.
Comprehension debt. Osmani's sharpest warning about his own idea: loops can ship work the author no longer understands. He calls the failure mode "cognitive surrender."
Here's the thing: I agree with all three. We didn't build Auteryn to argue with these objections. We built it because of them — starting months before the word existed.
What a loop needs, and what we run
Strip the buzzword and a trustworthy loop needs six things. This is what each one is on Auteryn, in production, today:
- A trigger that isn't you. Loops wake on schedules and on events — a webhook fires, a customer messages, a repository changes. That surface has been live since the platform's first release, under the unglamorous name of Flows.
- Wake-on-event, not polling. An agent can start a build in a background shell and keep working; when the job exits, the agent is told on its next turn. No token-burning "are we there yet" loops — the harness watches so the model doesn't have to.
- State that survives the gap between passes. Each agent works on a persistent machine — terminal, browser, filesystem — with snapshots, a working-memory scratchpad, and long-term memory with semantic recall. A pass picks up where the last one stopped.
- A clean landing, every time. Runs carry hard budgets of wall-clock time and model calls. Near the budget, the run doesn't die mid-thought — it checkpoints its todos and notes and lands. Interrupted runs resume with plan, tool history, and notes intact.
- Verification that isn't self-grading. Work gets checked by things that aren't the model that produced it: tests run in the sandbox, a vision agent that drives the real UI the way a human tester would, and a reflection pass after the run.
- Stop rules a human wrote. Every agent has an autonomy dial — Autopilot, Guarded, Supervised, Co-pilot — and approvals ride the loop: a scheduled pass obeys the same rules as one you started by hand. A run can park on a question for up to 72 hours and resume with full state when you answer.
Notice what that list is: it's the skeptic ledger, answered item by item. Unattended mistakes → verification plus stop rules. Cost → hard budgets, a clean landing, and credits you can actually read (1 credit = $0.001, forever — and free-plan runs are capped at 150 credits each, structurally). Comprehension debt → the part we're proudest of, and the next section.
The seat
Every loop-engineering diagram published since June has the same shape: goal in, loop in the middle, result out. A human appears twice — writing the goal at the start, reading the report at the end. The middle is a black box that emails you.
We think the middle is the product.
On Auteryn, one pass through the loop is a run, and a run is a place you can be: the live plan, the model's reasoning, and every tool call stream into a timeline as they happen; the agent's terminal is a tmux session you can type into alongside it; you can take the browser cursor, fix something, and hand control back. When other agents finish, they send you a replay. Here, you were in the loop — literally.
That's also our answer to comprehension debt. You don't lose understanding of work you watched — and occasionally steered — the way you lose understanding of work that happened to you.
The industry's verdict on agent loops is that they still need humans in them. We agree. We built the seat.
Where this goes
One roadmap note, labeled as such: we're building toward loops you create by saying them — "check the issues in my repo every hour" and the agent proposes the loop, shows you what it will cost, and starts only when you confirm. The pieces underneath it are the ones described above, and the confirmation-first, budget-first shape is the whole point. When it ships, it will appear on this page as a recorded run, not a promise — that's house policy for every claim on our site.
Until then: the loops that exist today — scheduled, event-driven, checkpointed, budgeted, with a seat in them — are live on the free plan, which is a real agent on a real machine, 1,000 credits a month.
Sources, as cited: The New Stack on Cherny and loop engineering; Peter Steinberger's June 7 post; Addy Osmani's "Loop Engineering"; The Register, June 24; the Hacker News thread. Third-party figures (token multiples, the Uber cap) are reported by those sources, not our measurements. Product claims map to the public claims register we keep in the repo.