Anthropic introduced a multi-agent harness design to support long-running autonomous application development for both front-end design and full-stack software creation. This approach divides tasks between separate agents responsible for planning, generation, and evaluation, with the goal of maintaining consistency and improving output quality across hours-long AI sessions.
This design addresses common challenges in autonomous coding workflows, such as context loss and premature task termination. Anthropic engineers implemented context reset along with structured handoff artifacts to allow the next agent in the workflow to continue from a defined state. This is different from context-preserving compression, but it can make the model wary of approaching context limits, which can impact performance on augmentation tasks.
Self-assessment of outcomes was also a focus. Agents often overestimate their results, especially for subjective tasks such as design. To alleviate this, Anthropic introduced a separate rating agent that was calibrated with a small number of shot examples and scoring criteria.
said Prithvi Rajasekaran, Head of Engineering at Anthropic Labs.
Separating the agents who do the work and the agents who judge it proves to be a powerful means of addressing this problem.
For front-end design, the team established four criteria: design quality, originality, craftsmanship, and functionality. Evaluators navigate the live page, interact with the interface using Playwright MCP, and provide detailed critiques to guide the generator through the iteration cycle. Each cycle produces progressively more refined output. The number of iterations ranges from 5 to 15 per run, which can take up to 4 hours in some cases, to produce designs that combine visual distinction with functional accuracy.
Industry experts emphasize the framework’s structured approach. Artem Bredykhin wrote on LinkedIn:
Long-running AI agents fail for a simple reason. All new context windows will be amnestic. The breakthrough structure is a JSON functional specification, forced tests, per-commit progress, and an init script that ensures every session starts with a working app.
Another observer, Raghu Arangarajan, commented on LinkedIn:
The three-agent framework provides a repeatable workflow for multi-hour sessions and ensures that evaluation and iteration are separated from production, improving overall reliability and output quality.
Anthropic engineers have applied this framework to various types of tasks and evaluated performance improvements. They observed that by separating planning, generation, and evaluation, subjective evaluations could be better handled while maintaining reproducibility of objective tasks. Structured multi-agent workflows also facilitate incremental progress in long-running sessions by clearly defining responsibilities and handoffs between agents.
Operationally, teams must establish evaluation criteria and adjust scoring mechanisms while monitoring iterative output. Although the agent performs the evaluation automatically, human oversight is still important for initial calibration and quality verification. This workflow supports distributed processing of tasks, allowing multiple agents to run in parallel or sequentially based on dependencies.
As AI models improve, the role of the harness may change, with some tasks being handled directly by next-generation models. At the same time, the improved model allows the harness to tackle more complex tasks. As the space of harness combinations evolves depending on the model’s capabilities, engineers must experiment, monitor traces, decompose tasks, and adjust harnesses.
