Breaking

Anthropic Details Three-Agent Harness for Long-Running AI App Development

September 17, 2026 at 04:18 EDT

AI Agents
Software Dev & Coding

Anthropic has described a multi-agent harness for long-running app development that splits work among a planner, a generator and an evaluator rather than relying on one autonomous coding agent. The company says the setup produced a working 2D game-maker prototype where a solo Claude Opus 4.5 run produced an app whose play mode did not work, but the stronger result came with far higher cost and runtime.

Mar. 24, 2026 · Anthropic

Planner, builder, reviewer

A long-running app harness split the job into clear roles: one agent scopes the product, one builds it, and one tests it like a user. The result worked better than a solo run, but took far more time and money.

agent roles

features in the game-maker spec

sprints in the full harness

$124.70

later DAW harness cost

The quality jump had a real price

Cost comparison for the 2D retro game-maker task.

Solo run

$200

Full harness

Scale: one full block is about $10; the solo run is shown as a proportional short block.

The handoff loop

Planner

Expands a short prompt into a product spec.

→

Generator

Builds features against an agreed contract.

→

Evaluator

Clicks through the app and sends actionable QA feedback.

What improved

The full harness produced a playable game-maker build where the solo run's core play mode failed.

What still matters

The better build cost more than 20x as much, took six hours, and still left rough edges for QA and users.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.