- Author
- Justin Young, Anthropic
- Published
- Nov 26, 2025
- Kind
- Company engineering blog post describing internal experiments
- Original
- www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
- Subject
- Harnesses and agent loops
- Themes
- Handing over control, What lasts
What it is
Justin Young writes for Anthropic about problems with agents that work for hours or days. The post sets out to give a setup that bridges coding sessions.
What it says
- Each new agent session starts with no memory of the last, which hampers long tasks.
- Anthropic's fix is an initializer agent first, then a coding agent each session.
- Early failures were doing too much at once, and declaring the job done early.
- The initializer writes a JSON feature list of over 200 items, all marked failing.
- Each session works on one feature, then commits to git and writes progress notes.
- Claude marked features done without end-to-end tests until prompted to use browser automation.
- The post leaves open whether one general agent or several specialized agents works better.
What evidence it gives
The evidence is internal experimentation on one task, a claude.ai clone. The post shows a feature-file format, a sample session and a failure-mode table. It reports no measured success rates.
Interests and limits
Anthropic wrote the post and makes the Claude Agent SDK it describes. Only one kind of project was tried: a full-stack web app. The post notes browser tools missed some bugs, such as native alert modals. It calls generalization to other fields a future direction.
Connections
These links are this wiki's own reading, not claims the sources make.
- A spec the agent can't run isn't a spec yet. Young's team chose JSON because models overwrite it less often than Markdown. It added strong wording against editing tests. Lewis says anything written mid-task grades itself generously. So file format has become a soft lock. How firm must the definition of done be?
- Prompts are code, .json/.md files are state. Zechner and Young's team separately reach for JSON on disk. Each new session starts cold, so progress must live in a file. Zechner's plan comes from a script, and Young's from an agent. Which author of the plan can you trust more?
- Upaya. Upaya adapts a teaching to its audience, and the method need not be ultimately true. Young's team tells a model that removing tests is unacceptable. That sentence is chosen for its effect on the reader. Prompts may be the newest expedient means.
- Assessors of Maat. Book of the Dead chapter 125 pairs each of 42 judges with a matching declaration. Young's feature list has over 200 items, all marked failing at first. Both are checklists for a verdict. The Egyptian scheme still weighs the heart. Young's agents marked items done until browser tests arrived.