Google researchers have open‑sourced EnvHarness, a tool that automatically reshapes an agent’s training world to target its weak spots. It sits on top of any existing environment that uses the standard reset() and step() calls, adding plug‑in components that change start states, rules, or observations without touching the core task or its verifier. The built‑in EnvRigger watches the agent, finds recurring failures, creates focused mutations, and validates them through new interactions. Tests showed models like Gemini 3.5 Flash and Claude Sonnet 4.6 improve by 2.9–3.7 percentage points on benchmarks such as WebArena and ALFWorld.
Why it matters
Developers can train stronger agents faster without rebuilding benchmarks or verifiers.
