Google open-sources Harness², turning any agent harness into a self-improving system
rez0__ · x · 2026-10-20
Google Research open-sourced Harness², which turns any harness — the system prompt, tools, memory, skills and subagents wrapped around a model — into a self-improving system.
How it works: model weights stay untouched. On each new task the agent proposes edits to its own harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved. Because the learning signal comes from run-to-run differences, it needs no ground-truth answers or grader.
Early adopters plugged it into JEV 1.14 and report large gains on vision benchmarks across rounds with zero weight changes — the system rewrites its own prompts, builds its own tools, and distills its own experience. The JEV-Harness² repo is coming soon.
More from coding & agent
- Perplexity's Open-Source Sandbox Platform SPACE Joins Crusoe Cloud's Managed AI Portfolio — AravSrinivas · 2026-10-20
- SpudStack uses Grok to review agent plans with PASS/FAIL verdicts — agents can't sign off their own work — alexcovo_eth · 2026-10-20
- Cloudflare MCP Server adds step-up auth: agents request scopes on demand — irvinebroque · 2026-10-20
- Skill Harbor, a skill directory for Muse agents, opens with a 10-member ecosystem — InformationOk4877 · 2026-10-20
- Nine conditions for AI to actually show up positive in your P&L — vasuman · 2026-10-20
- Three probes track post-training SWE-bench rankings, BPB hits ρ=0.964 — hyunw_kim · 2026-10-20