독서리서치레이더투자 프레임워크
로그인 / 회원가입
← 읽기로 돌아가기
뉴스간단한 요점1 분 소요DisconfirmAI 편집부

Opus 5.5 stops midway, and Anthropic says old agent loops are misreading it as done

Anthropic's own guide confirms the main cause: programs treat a progress report as completion. Developers running agents should check their continuation loops first.

2026-09-26 22:20 GMT+8

Anthropic's official guide confirms a pattern reported by developers: unattended Opus 5.5 agents stop midway not because the model is lazy, but because old agent loops treat a plain-text progress report ending in end_turn as task completion.

The guide names four stopping habits: summarizing without acting, pausing politely for permission, listing decision items that do not block progress, and stopping to summarize after each small stage. The fixes it proposes: a task list with automatic continuation, a smaller model checking output against a completion standard, and handing the task to a human after two or three stalled retries.

The same migration guide lists four changes that make old code fail with a 400 error: thinking can no longer be disabled, tool_choice can no longer force a tool, thinking blocks are bound to model and context, and the legacy computer-use tool is retired on the Claude API and Google Cloud. Per IT之家's report on the guide, thinking content is not returned by default yet is still billed as output tokens.

출처

  • ithome.com↗
리서치 계속하기 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크