Harness design for long-running application development
01:00 · March 24, 2026 · Antropic Engineering Blog
Harness design is key to performance at the frontier of agentic coding. Here's how we pushed Claude further in frontend design and long-running autonomous software engineering.
Summary
Anthropic engineer Prithvi Rajasekaran describes iterative work on multi-agent harnesses that enable Claude to produce complete frontend designs and full-stack applications over multi-hour autonomous runs. The approach draws from generative adversarial networks by pairing a generator agent with a separate evaluator, then adds a planner agent to expand brief user prompts into detailed product specifications. This structure addresses two recurring limitations: progressive loss of coherence as context windows fill, and the tendency of models to overrate their own outputs, especially on subjective criteria such as visual design.
Earlier harnesses relied on an initializer that produced task lists and a coding agent that advanced one feature at a time, passing structured artifacts between sessions. Persistent failures included “context anxiety,” in which models prematurely concluded work near perceived token limits, and lenient self-assessment that allowed mediocre implementations to persist. Context resets combined with explicit handoff artifacts restored coherence, while isolating evaluation in a dedicated agent made reliable critique feasible. The evaluator received Playwright access to interact with live pages or running applications, scoring outputs against explicit criteria rather than relying on the generator’s judgment.
For frontend tasks the criteria weighted design quality and originality above basic functionality, steering generations away from generic layouts toward more distinctive solutions. With Claude Opus 4.5 the model’s reduced context anxiety allowed the team to drop explicit resets in favor of continuous sessions managed by automatic compaction. The resulting three-agent system produced richer applications while trading increased orchestration overhead and longer wall-clock time for measurable gains in scope and quality.
Why it matters
This article provides highly actionable architectural patterns for product teams and builders developing autonomous AI agents. It offers concrete solutions to common LLM limitations like context degradation and self-evaluation bias, which are critical for Dutch AI engineering teams building robust, long-running applications.
