EnvHarness: co-evolving environments push agents from 47.67% to 54.79% on SWE-bench

青稞AI · wechat · 2026-08-22

Google researchers propose EnvHarness, an "environment scaffold": instead of only letting agents explore static environments, give environments dynamically modifiable scaffolds so agent and environment co-evolve.

Three components: Stage reshapes initial states via hidden preset actions after reset; Contract takes over environment internals, rewriting action spaces, transition dynamics, and observations; Chain stitches environments together or branches on intermediate results for longer-horizon tasks.

An "EnvRigger" agent automates scaffold construction via a four-step loop — Observe (collect trajectories), Diagnose (find failure modes like shortcut exploits), Write (synthesize components), Validate (verify the new environment teaches new skills and stays solvable, else rewrite).

On SWE-bench Verified, agent success rises from 47.67% to 54.79% as custom environments scale, still climbing at 300 environments. Future work: richer components (randomness, partial observability, multi-agent), non-text (visual/GUI/embodied) environments, and semantic task composition.

(Paper: arxiv.org/abs/2608.19880, code open-sourced)

Original post →

EnvHarness: co-evolving environments push agents from 47.67% to 54.79% on SWE-bench

More from coding & agent