A new training technique lets AI agents rehearse tool use internally rather than relying on external environments, according to a paper published on arXiv introducing EnvACE. The method, which the authors call "world rehearsal," allows a single model to generate a tool call and then simulate the response that call would trigger, with both functions optimized together using task-success rewards. Across four benchmarks—BFCL-v4, tau²-Bench, VitaBench, and FinMCP-Bench—the authors report that EnvACE outperforms baselines that scale external environments during reinforcement learning training.
The world rehearsal mechanism works through what the authors describe as a role swap inside one model. The policy produces a tool call, then assumes the role of the environment to generate the response that action would have caused, and finally uses that rehearsed response to inform its next move. Both the action-selection weights and the environment-prediction weights are trained end-to-end with the same task-success signal, meaning the parameters that choose what to do also learn what each tool would return. The paper also describes a test-time behavior where the internalized world model lets the agent rehearse a plan privately before executing a real tool call, which the authors say produces additional performance gains under a moderate rehearsal budget.
The authors write that the recipe for training capable tool-use agents has been growing more expensive rather than less expensive, forcing teams to either connect agents to real executable environments and bear the cost of building and verifying them, or deploy synthetic simulators that are difficult to ground. According to the report, the claim is that EnvACE beats environment-scaling baselines in overall evaluation across the four benchmarks, though the paper does not break out margins for each benchmark, explain how the internalized environment handles real API drift after training, or show what happens with tool patterns the model hasn't seen. The report advises readers to take the specifics as reported rather than settled.
The compute and engineering cost of environment simulators has been a quieter bottleneck in agent training, the report notes. If a portion of that infrastructure can shift into the model's own parameters, the barrier to training more capable tool-use agents falls for teams that can't afford to maintain large simulator or sandbox infrastructure. The report suggests that if world rehearsal withstands further scrutiny, the next generation of agent training runs may resemble teaching a single model to function as its own training gym rather than scaling external environments—a fundamentally different cost curve for anyone building agents outside the largest labs. For organizations betting on agentic workflows, the choice between investing in environment infrastructure versus model capacity just became less binary. If internalized simulation proves durable, competitive advantage in tool-use AI may tilt further toward teams that can architect training loops efficiently, not just those with the budget to simulate the world at scale.