Google Research Proposes EnvHarness to Improve Agent Training Signals
机器之心 · wechat · 2026-08-23
Google Research has released a paper titled 'EnvHarness: Awakening Static Worlds for Agent Learning' along with an open-source project, introducing a new component called EnvHarness. Analogous to adding harnesses (e.g., memory, skills) to LLMs, EnvHarness operates on the environment side. It controls how the environment interacts with the agent via standard interfaces without altering the underlying environment (e.g., web pages, codebases, robotic scenes) or the agent's core model.
The paper defines three types of EnvHarness components:
- Setup: Changes the environment state after initialization.
- Interaction Rule: Maps agent actions and observations to new ones.
- Linking: Switches to another environment upon meeting certain conditions.
The method uses a 'loop engineering' process where an agent observes trajectories, identifies issues, and helps generate or optimize the EnvHarness. This cycle ensures the environment provides better training signals. Experiments show that EnvHarness effectively improves performance in reinforcement learning, self-evolving agents, and across benchmarks like embodied intelligence, web navigation, and code generation.
Related event: Google's EnvHarness Makes Environments Co-Evolve with Agents(6 posts)→
More from coding & agent
- EvoOntology: a self-evolving ontology layer bridges the agent-data gap — RUC-DataLab · 2026-09-21
- Xiaomi's CodeMidas builds 5,545 coding RL environments from raw source code — XiaomiMiMo · 2026-09-21
- GraphSkillEvo evolves graph-structured skills for LLM agents, +4% on benchmarks — Rui Sun · 2026-09-21
- Dev shares multi-model workflow: Grok for daily coding, GPT-6 for hard bugs — minchoi · 2026-09-21
- Meta Muse pitch revives the question: will personal AI agents get real permission controls or just one big Allow button? — yi111 · 2026-09-21
- A Monday-ready checklist for decision models: calibration, cost-based thresholds, pinned versions — colinmcnamara · 2026-09-21