Agent training moves off costly live environments

Organizations

Updates

1. Aug 21

   Google's EnvHarness adds a plug-in layer that reshapes static environments without changing their logic or verifier, and its automated rigger targets a policy's diagnosed flaws for up to 9.0-point gains with 9.8% fewer execution steps.

   Google's EnvHarness reshapes agent training environments

2. Aug 7

   EnvACE replaces executable environments with a policy that also simulates the environment's responses, and AgentOPSD reaches 89.1% on ALFWorld without extra critics or rollouts; Google's reported ~$1.5B talks for Mechanize's environments show live environments still command a premium.

   EnvACE trains agents through world rehearsal · Tsinghua's AgentOPSD critic-free turn-level credit

   Against Google in talks to license Mechanize for ~$1.5B

3. Jul 24New

   Experience Distillation kept at least 64.8% of in-context gains in weights and matched RL with at least 9.6x fewer environment samples, LLM-as-a-Coach distilled judge feedback to beat rubric-based RL on non-verifiable tasks, and Apple synthesized API-agent trajectories with no executable backend.

   Experience Distillation internalizes agent trial and error · Microsoft Research's LLM-as-a-Coach · Apple synthesizes API-agent data without environments

Predictions

1. Jan 2027

   Distilling an agent's own experience, rather than running expensive RL rollouts, becomes a default post-training stage and lowers the environment cost of building reliable agents. Jul 24

2. Feb 2027

   Self-generated supervision and internalized world models become standard components of agent post-training, cutting the environment and rollout cost of building reliable agents. Aug 7

3. Feb 2027

   Environment- and harness-level optimization plus inference-serving kernels become a distinct layer of the agent stack, decoupling capability gains from fresh pretraining. Aug 21