By Leela Kumili
Publication Date: 2026-04-04 14:24:00
Anthropic has introduced a multi-agent harness design to support long-term autonomous application development, targeting both front-end design and full-stack software creation. The approach distributes tasks across different agents responsible for planning, generation and evaluation, with the aim of maintaining coherence and improving output quality over multi-hour AI sessions.
The design addresses challenges common to autonomous coding workflows, such as: B. Loss of context and premature task termination. In addition to structured handover artifacts, Anthropic engineers also implemented context resets so that the next agent in the workflow can continue from a defined state. This is different from compression, which preserves context but can make models cautious as they approach context boundaries, potentially affecting performance on advanced tasks.
Another focus was self-assessment of the results. Agents often overestimate their results, especially for subjective tasks like design. To mitigate this, Anthropic…
