How Building Software Is Changing at Anthropic
- Source
- newsletter.pragmaticengineer.com
- Author
- Gergely Orosz
- Date
Context
Faster prototyping still left Anthropic with substantial architecture work when building Claude Managed Agents, its infrastructure for running agents. In Gergely Orosz's reporting from the company, the team describes a roughly six-month project that used an internal customer to test its design, then changed that design midway through development.
The resulting separation explains why the redesign mattered. The 's control loop, execution environment and event history became independently replaceable components. As the company's engineering account explains, a failed control loop can restart from the stored event log instead of losing the session with its container. That architecture background shows the reliability work behind the product, beyond generating its code.
- The rearchitecture that made Managed Agents work: split the 'brain' (model plus ) from the 'hands' (sandboxes and tools) and the 'session' (the event log), each an interface that assumes almost nothing about the others.
- Planning, not coding, was the biggest phase of a six-month agent-infra project, and the artifact was still a Google Doc PRD — because its job is aligning the team, cloud partners, and product, which AI does not speed up.
- They de-risked the product by building it for an internal customer first: the backend for Claude Code on the web was the same shape of problem, so it served as the spike before generalizing it for outside customers.
- Negotiating interfaces between teams got cheaper. Instead of handing over a fully spec'd interface, they stood up a stub service that shadowed traffic and settled the contract while both sides kept building.
- Anthropic's API layer is migrating from Python to Rust. Python was picked for researcher familiarity and fast iteration, but it is single-threaded and does not hold up on the hot path under load.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
“Historically, rewrites are a terrible idea.”
Jarred Sumner
“We see that few tokens are spent on actual implementation. Most are spent on discovery of unknowns, prototyping, mocking, and then in verification and testing.”
Thariq Shihipar
Checking sign-in…
Loading comments…