AI engineering has left the prompt behind and started writing its own instructions
Eight Sleep's engineers stopped writing code a year ago. Anthropic now ships a feature where AI generates its own prompts. The synchronous, human-initiated model of AI-assisted work is giving way to something that runs whether or not anyone is at the keyboard.
Matteo Franceschetti, the chief executive of Eight Sleep, puts the transition in terms that resist softening. His engineering team stopped writing code roughly a year ago. In place of human engineers writing code, the company runs hundreds of AI engineers that code on their behalf, and internally operates hundreds to thousands of agents. Each morning, the marketing and growth team receives a report from those agents suggesting what changes to make. The team approves, moves, or rejects. Then the agents proceed.
That is not a description of productivity tooling layered onto an existing workflow. It is a different organizational structure, with AI occupying the layer where junior and mid-level execution used to live.
Fred Turner describes the same architecture from an engineering management angle. In his account, one senior engineer manages a set of downstream agents that do the actual implementation work. The agents take feedback from the team on features, build them out, and the engineer comes in afterward to verify that what was built makes sense. The human role has shifted from writing to reviewing. Turner’s framing and Franceschetti’s operational report describe the same underlying model from different vantage points.
Our engineers stopped coding around a year ago. So since then they didn't code. what they have is hundreds of AI engineers that they code for themMatteo Franceschetti
Fiona Fung, who works at Anthropic, adds a structural detail that clarifies why this shift is accelerating. Anthropic recently launched what she calls routines, a feature where the AI generates prompts autonomously rather than waiting for a human to initiate each exchange. The significance is the abstraction level. Previously, a human had to prompt. Now the system prompts itself, triggered by conditions or schedules rather than by a person sitting at a keyboard.
Jason Lemkin reports that routing workflows across multiple models has become standard practice across all but the smallest startups, and that this shift happened within roughly the last 90 days. Martin Casado, a general partner at Andreessen Horowitz, offers one reason the pattern is spreading: running the same query through multiple models adversarially and using a separate model to reconcile the results, he argues, yields more information than any single model produces alone. Greg Isenberg describes the practical architecture of these loops in simpler terms: code, a thinking loop, and a live data stream. He also describes the control mechanism his team uses, getting pinged on Slack each time a loop completes its run, a lightweight human checkpoint that keeps the system from operating entirely without oversight.
The scale available to those willing to build at this layer is not modest. Greg Brockman, president of OpenAI, says his organization deployed 10,000 agents to solve the Navier-Stokes equations. He also describes an AI system that ran coherently for 24 hours on long-horizon tasks across a wide variety of domains. Those figures mark a ceiling that synchronous, prompt-by-prompt interaction cannot approach. Brockman locates the binding constraint for agentic systems not in model capability but in tools and context: whether the model is smart enough to use the tools available and whether it has access to the context it needs. The rapid spread of Model Context Protocol servers and command-line interfaces, which Brockman describes as taking the world of software and making it accessible in ways not designed for human interaction, is what is closing that gap.
Harry Stebbings offers a more granular unit of measurement. His two-and-a-half-person team runs Replit between 10 and 20 hours a day. At the end of last year, they were running it about an hour a day. The difference is not a new strategy. It is the same tool, operated at a volume that only became viable as the underlying models improved. Lemkin projects that within a year, model and harness improvements will eliminate the need for many of the manual corrections that routing workflows currently require. The infrastructure, in other words, is still being built while it is already in use. The evidence on the ground points in one direction: self-prompting systems are already treating the human as a checkpoint rather than an initiator. The question for any engineering organization still operating synchronously is not whether this model is viable. It is how far behind synchronous operation leaves them.