Anthropic's Three-Agent Harness: Why Conservative Planning Produces Underwhelming Results

Anthropic published something genuinely useful on April 4. Not a model announcement, not a benchmark claim. An engineering blog post detailing how they build production apps with a three-agent harness: Planner, Generator, Evaluator.

The architecture is straightforward. The Planner takes your prompt and expands it into a detailed spec. The Generator builds in sprints. The Evaluator uses Playwright MCP to actually click through the running app and score it against four criteria: design, originality, craft, functionality. A solo agent run costs about $9 for 20 minutes of work. The full three-agent harness runs $200 over 6 hours but produces production-grade full-stack applications.

The non-obvious finding is what matters here. Conservative planning consistently produces underwhelming results. The Planner must be deliberately ambitious. I've seen this in my own work. When I give Claude Code a cautious, well-scoped prompt, I get cautious, well-scoped output. When I push it to think bigger, describe the full vision, demand more, the output quality jumps.

This converges with two other findings this week. Addy Osmani's "Code Agent Orchestra" framework says three focused agents consistently outperform one generalist working 3x longer. LangChain's context engineering paper formalizes the four strategies (Write, Select, Compress, Isolate) that make multi-agent architectures actually work. The industry is converging on the same conclusion: the harness around the model matters more than the model itself.

For builders: if you're still prompting a single agent in a single session for complex work, you're leaving quality on the table. The Planner-Generator-Evaluator pattern is simple enough to implement this weekend. Start ambitious. Let the Evaluator be the one that reins things in.