From Startup Fortune · · 1m 20s
Google Releases Harness2 for AI Agent Self-Improvement
Google's Harness2 lets AI agents rewrite their own playbook without model changes.
In brief
Google's Harness2 lets AI agents rewrite their own playbook without model changes. Google open-sources Harness2, an AI framework that enables agents to self-improve by editing their own scaffolding without altering model weights. Originally reported by Startup Fortune.
Google Cloud AI Research releases Harness2
Google Cloud AI Research released Harness2 on October 9 under an Apache 2.0 license to let AI agents rewrite their own scaffolding without manual patches.
Harness2 doesn't change model weights
Harness2 keeps model weights frozen while editing instructions, tools, memory, and routines that decide agent actions.
Propose-Probe-Compose mechanism
The Propose-Probe-Compose framework proposes two edits, probes them, and composes changes based on observed behavioral contrast without labeled correctness.
Performance gains on benchmarks
Harness2 added 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB before recursion.
Claude Opus as editor model
Google used Claude Opus 4.8 to propose and compose harness changes, a detail highlighting Anthropic's role in the framework.
Self-improvement debate intensifies
Harness2 makes harness engineering the most plausible near-term path for agent self-improvement, shifting focus from model weight updates.
Limitations of self-improvement
Systems often stall after one structural fix due to lack of internal model for why changes worked, leading to reward hacking without compounding gains.
The short version
Google open-sources Harness2, an AI framework that enables agents to self-improve by editing their own scaffolding without altering model weights.
