Google Open-Sources Harness², Letting Agents Rewrite Their Own Scaffolding
A newly released framework turns any agent harness into a self-improving system, with the agent proposing edits to its own prompts, tools, and memory after every task — no human grader required.
Google has quietly open-sourced Harness², a framework that lets AI agents modify the very scaffolding they run on. According to a widely-shared thread from @akshay_pachaar, the core idea is deceptively simple: on every task, the agent proposes edits to its own harness — the system prompt, the tool definitions, the memory structures — and those edits persist into future runs. The thread drew roughly 4,758 likes and 369,000 views over the weekend, a level of attention that reflects how directly it touches a problem working agent builders keep hitting.
The significance is architectural. Most production agents today run on static scaffolding: a human writes the system prompt, wires up the tools, defines the memory schema, and then the agent operates inside that fixed envelope. When the agent fails in a predictable way, a human notices, diagnoses, and patches the harness by hand. Harness² collapses that loop. The agent becomes responsible for its own iteration, treating the scaffolding itself as something to be optimized rather than a fixed given.
Get our free daily newsletter
Get this article free — plus the lead story every day — delivered to your inbox.
Want every article and the full archive? Upgrade anytime.
No spam. Unsubscribe anytime.