Anthropic Published How They Actually Build With Claude Code. It's Two Agents, Not One.

The most useful engineering blog post I've read this year dropped today with zero fanfare.

Anthropic's engineering team published the actual architecture they use for long-running autonomous coding: a two-agent harness where an initializer agent sets up the project environment (init scripts, feature lists, progress logs) and a coding agent works incrementally across multiple context windows. The memory bridge between them? A plain text file called claude-progress.txt. Not a vector database. Not a knowledge graph. A text file.

The specifics are what make this valuable. They expand prompts into 200+ discrete end-to-end features tracked in a JSON structure. Each feature has a defined scope, acceptance criteria, and verification method. Puppeteer-based browser testing runs automatically to verify UI features. The coding agent reads the progress file at the start of each context window, picks up where the previous session left off, and updates the file when done. It's embarrassingly simple and that's exactly why it works.

This post generated 3,460 likes and 435K views. That engagement isn't because people found it interesting. It's because people found it useful. I've already started adapting the pattern for my own projects. The progress file approach solves the most painful problem with long-running agent sessions: context loss between windows. I've been doing something similar with CLAUDE.md files, but the structured JSON feature tracking is a level up.

The broader signal: "harness engineering" is becoming a real discipline. The agent itself is a commodity. The harness around it, the scaffolding that manages context, tracks progress, verifies outputs, that's where the craft is. If you're building anything with autonomous coding agents, read this post and steal the architecture. It's production-tested by the team that built the model.