Anthropic has introduced a multi-agent harness design to support long-running autonomous application development, targeting both frontend design and full-stack software creation. The approach divides tasks among distinct agents responsible for planning, generation, and evaluation, aiming to maintain coherence and improve output quality over multi-hour AI sessions.
The design addresses challenges common in autonomous coding workflows, such as context loss and premature task termination. Engineers at Anthropic implemented context resets alongside structured handoff artifacts, enabling the next agent in the workflow to continue from a defined state. This differs from compaction, which preserves context but can make models cautious about approaching context limits, potentially affecting performance in extended tasks.
Self-evaluation of outputs was another focus. Agents often overrate their results, particularly on subjective tasks such as design. To mitigate this, Anthropic introduced a separate evaluator agent calibrated with few-shot examples and scoring criteria.
Prithvi Rajasekaran, engineering lead at Anthropic Labs, says
Separating the agent doing the work from the agent judging it proves to be a strong lever to address this issue.
For frontend design, the team established four grading criteria: design quality, originality, craft, and functionality. The evaluator navigates live pages, interacts with the interface using Playwright MCP, and provides detailed critiques to guide the generator in iterative cycles. Each cycle produces progressively refined outputs. Iterations range from five to fifteen per run, sometimes taking up to four hours, and generate designs that combine visual distinction with functional accuracy.
Industry practitioners have highlighted the framework’s structured approach. Artem Bredikhin wrote on LinkedIn that
long-running AI agents fail for a simple reason: every new context window is amnesia. The breakthrough is structure: JSON feature specs, enforced testing, commit-by-commit progress, and an init script that ensures every session starts with a working app.
Another observer, Raghus Arangarajan, commented on LinkedIn that
The three-agent framework provides a repeatable workflow for multi-hour sessions and ensures that evaluation and iteration are separated from generation, improving overall reliability and output quality.
Anthropic engineers applied this framework across different task types to assess performance improvements. They observed that separating planning, generation, and evaluation enables better handling of subjective assessments while maintaining reproducibility in objective tasks. The structured multi-agent workflow also facilitates incremental progress in long-running sessions by clearly defining responsibilities and handoffs between agents.
Operationally, teams must establish evaluation criteria and calibrate scoring mechanisms while monitoring iterative output. Agents execute evaluations automatically, but human oversight remains important for initial calibration and quality validation. The workflow supports distributed processing of tasks and allows multiple agents to run in parallel or sequentially based on dependencies.
As AI models improve, the harness’s role may shift, with some tasks handled directly by next-generation models. At the same time, improved models enable harnesses to tackle more complex work. Engineers should experiment, monitor traces, decompose tasks, and adjust harnesses, as the space of harness combinations evolves with model capabilities.