The Agentic Engineer
- Anthropic published the blueprint for multi-hour coding agents. Three-agent harness (planner, generator, evaluator) with hard context resets and GAN-style feedback loops. The most detailed production architecture any frontier lab has shared.
- GitHub shipped /fleet for parallel multi-agent coding and open-sourced the Copilot SDK in 5 languages. The orchestration layer war is heating up.
- Amazon Nova Act MCP gives your agent a browser with one install. Natural language web automation, parallel sessions, built-in bot detection. Stop writing brittle Playwright scripts.
Anthropic's Three-Agent Harness: The Blueprint for Multi-Hour Autonomous Coding
Single-agent coding sessions hit a wall around 30 minutes. Context fills up, the model starts hallucinating file paths, and quality degrades. Anthropic just published the architecture they use internally to keep coding agents productive for 4+ hours. It's three agents, hard context resets, and a feedback loop borrowed from GANs.
The setup: a planner decomposes your spec into chunks that fit a single context window. A generator builds code for each chunk. An evaluator grades the output by actually running it, navigating the live UI with Playwright MCP, and checking results against the spec. If the evaluator fails the output, the generator gets specific feedback and tries again. 5 to 15 iterations per run.
Read the full issue
Subscribe to The Agentic Engineer to unlock this and every future issue.