How Anthropic’s Multi‑Agent Harness Keeps Claude Running for Hours

Introduction

Anthropic’s recent blog post analyzes why large‑language‑model agents like Claude tend to truncate long tasks ("context anxiety") and over‑estimate their own outputs ("self‑evaluation distortion"). Their solution is a multi‑agent harness that isolates generation, evaluation, and planning, similar to a GAN architecture.

Key Problems

1. Context Anxiety

When the model senses its context window filling up, it becomes conservative, compresses exploration, and rushes to finish, often delivering sub‑optimal results. Anthropic mitigates this by context reset : after each stage, a new agent receives a structured hand‑off file instead of the full, near‑full context.

2. Self‑Evaluation Distortion

Claude struggles to critically assess its own code or designs, frequently giving overly optimistic pass/fail judgments. The remedy is to fully separate generation from evaluation , assigning a dedicated evaluator agent to critique and request revisions.

GAN‑Inspired Multi‑Agent Design

The framework mirrors GAN’s generator‑discriminator split, adding a planner:

Generator agent : performs the actual work—writes code, creates UI, advances the task.

Evaluator agent : scores outputs, finds bugs, and suggests fixes.

Planner agent : expands high‑level requirements into detailed specifications without dictating implementation details.

This loop iterates: the generator produces a version, the evaluator critiques it, and the generator refines the work based on feedback.

Experiment 1: Front‑End Design

Anthropic tested the harness on a front‑end design task, exposing the typical “safe but bland” output of a single agent. They defined four scoring dimensions—design quality, originality, craftsmanship, and functionality—giving higher weight to design quality and originality.

The evaluator used Playwright MCP to interact with the live page rather than static screenshots, running 5‑15 iterations per round (each round up to 4 hours). After ten rounds, the generator shifted from a conventional dark theme to a 3D museum‑style experience, demonstrating that sustained evaluator pressure can drive creative breakthroughs.

Experiment 2: Full‑Stack Application (RetroForge)

Using the same three‑agent system, Anthropic built RetroForge , a 2D retro‑game creator. Compared to a single‑agent version (20 minutes, $9), the three‑agent harness took 6 hours and $200 but produced a fully playable product with functional gameplay, AI‑driven content generation, and proper QA.

During Sprint 3, the evaluator checked 27 acceptance criteria and reported concrete bugs such as:

Rectangle fill tool not triggered: fillRectangle missing call in mouseUp.

Entity spawn point deletion failure: missing selection flag when selectedEntityId is set.

Animation frame reorder API error: PUT /frames/reorder mis‑parsed frame_id as integer, causing 422 response.

These detailed findings show the evaluator acting like a strict automated QA, feeding precise fixes back to the generator.

Experiment 3: DAW Prototype with Opus 4.6

After upgrading to Opus 4.6, Anthropic observed reduced context anxiety, allowing the three‑agent system to run longer without explicit resets. They built a browser‑based digital audio workstation (DAW) in ~3 hours 50 minutes for $124.70, achieving a functional arrangement view, mixer, and transport, plus AI‑driven tempo and melody controls.

Limitations remained: Claude cannot hear audio, so the evaluator could only verify functional and UI aspects, not sound quality.

Engineering Takeaways

1. Evaluator Value

Evaluators are most beneficial for tasks near the generator’s capability boundary—complex, detail‑heavy projects where the model can produce a draft but may miss subtle issues.

2. Harness Components Reflect Model Limits

Each component (planner, generator, evaluator, context reset) encodes an assumption about the model’s shortcomings. As models improve, some components may become unnecessary.

3. Prompt Language Shapes Architecture

Evaluation criteria and prompt phrasing act as soft‑architectural constraints, directly influencing the generated product’s direction.

Future Directions

The community is discussing decentralizing the agents—making planner, generator, and evaluator independent services that discover and negotiate via open protocols, moving from a single runtime to a true agent network.

Conclusion

Anthropic’s multi‑agent harness demonstrates that separating generation and evaluation, combined with careful prompting and context management, can extend Claude’s effective runtime from minutes to hours, producing higher‑quality, more creative outputs across diverse domains.