Loop Engineering: how to design loops that run coding agents themselves

The conversation opens with the status of the source. The eleven pages look like a paper from a peer-reviewed journal, but, as Alexander puts it, this is a working note by an Anthropic employee published under his own name: no peer review, no preprint archive, no benchmark, just a narrative synthesis by an author whose name the captions render as Huashu — and Alexander admits he cannot tell whether that is a first name or a surname. Roughly half the text retells a short note by Addy Osmani published around 7 June. Maxim allows that the document did not leak by accident; Alexander adds that inside sits a ready example of a loop on Claude Code, and Maxim points to the chunk of markdown at the end you only have to copy into a tool: how else do you get people to work through a tutorial by hand? The structure, though, is good, so the discussion follows it.

The participants then reconstruct the week in early June when the concept got its name. Addy Osmani is a Google man who blogs and contributes to Google's white papers, including the May one on changes to the software development life cycle. He referred to Peter Steinberg, author of a product that took off completely unexpectedly last autumn; by February Steinberg had signed a contract with OpenAI, though by all accounts offers were plentiful. His thesis fitted into a single post: do not write prompts, construct loops in which agents compose their own prompts and watch their own exit conditions. At about the same time an expert from Anthropic who works on Claude Code showed tools that start an agent session on a schedule and let you state goals as exit conditions. That is how the ladder of abstraction lined up: prompts, context, harness — and the loop on top.

Alexander adds a framing of his own: the dumbness of a model used to be compensated by the complexity of the harness. In a book from last year by a Google employee, routing and the generator-critic pattern still had to be coded by hand on LangChain and ADK, and the agents came out limited; now the model writes the plan, hands out the work and reviews it with sub-agents, so the loop can finally close with the human out of it. Maxim adds that agent skills took off at the turn of the year: the difference between agents is no longer the model or the logic, but the skills loaded into it. The loop itself breaks into five steps: discovery, where work comes from; handing that work over so an agent can run it autonomously; verification able to say no, this is not good enough; persistence, the state between runs; and scheduling, when the loop wakes up next. Six components hold them up: skills, worktrees, memory, connectors, sub-agents playing generator and critic, and automations — in Osmani's version, as Maxim recalls it, five plus one, with memory pulled out separately.

Alexander explains persistence through the film 50 First Dates: the hero gives the girl with amnesia a diary, and every morning she loads a summary of her earlier life. Scheduling is ordinary cron, and viewers had asked whether the loop runs daily or hourly. Maxim starts with automations: coherent business scenarios for models are hard to invent, yet every company has plenty of events and errors, and the old Human by Exception pattern comes back to life here, with an agent taking the human's seat. Worktrees are about parallelism: a morning triage turns up ten to twenty tickets, running them sequentially makes no sense, but memory between sub-agents has to be shared carefully. Skills he calls programming for non-programmers: the behaviour of a universal agent is described in plain markdown, plus templates, rules and scripts in Python or PowerShell. To the list he adds the command line — in his view the most effective way to deal with the outside world, even for an API call.

Alexander answers with a story: the smartest models could not build this episode's deck through the MCP server of an online whiteboard, while the same deck as code came together over a few iterations — code wins, because models met Bash and Python examples in pre-training and never met the description of that particular MCP service: the hands are there, the taste is not. He compares verification to generative adversarial networks: work used to be hard to do and an experienced pair of eyes could judge it, now it is the other way round and difficulty has moved to acceptance. Verifying means running the product, going through probes and tests and collecting an evidence package. In his open source project Sonar keeps complaining about a parameter in a test that is a deliberately fake password, and Claude, Cursor and Codex react differently: one does not care about tests, another will fix them too. Skipping steps produces recognisable failures: without discovery the loop is blind; sloppy handoff once left Alexander with tangled pull requests on his system design book, where terminology fixes and reader feedback diverged and stopped merging; and without verification agentic development degenerates into vibe coding — he showed colleagues more deterministic checks than he had ever written in his life.

Maxim lists the four debts that pile up unnoticed: verification debt is invisible from outside, then understanding degrades, then comes cognitive surrender, when digging in stops being worth it, and finally token overspend. Investing in the harness, by contrast, saves tokens but shifts the cognitive load onto the human: Alexander recalls a talk by Matt Pocock on skills the model invokes and skills the user invokes — the more there are, the harder it is for agent and person alike to pick one. Maxim brings up the hierarchy of loops — generation, the agent loop, the loop with evals, a meta-loop: the higher you climb, the costlier the mistake, and the worst outcome is building the wrong thing well. Hence the boundaries: a session dies on its own, an automation lives on the half-closed laptop you carry around the office, while in the cloud security teams control the image and network access — at Alexander's company autonomous agents are moving into a sandbox, because drifted intent on a personal machine has a large blast radius. Both end on the same note: keep the first loop small — Maxim points to a dice-rolling skill that is easy to wrap in a loop with an exit condition — and the loop must run without a human yet stay stoppable.