DynaWeave turns static tool tasks into changing scenarios

When an agent uses tools to complete a task, a new request, unavailable service, or other outside change can make its original plan fail. Many task environments remain mostly unchanged after work begins.

Can existing tasks be turned into realistic changing scenarios, with reliable checks for whether the agent adapted to the changes that actually happened?

DynaWeave converts existing tool tasks into verifiable scenarios with changing conditions, then uses checked interaction demonstrations to train agents to adapt.

The authors used DynaWeave to create 2,095 verified interaction trajectories across 36 environments, then trained Qwen3.5 models further.

They compared models with and without this training on Gaia2, which tests changing applications; CostBench, which tests travel planning as costs and availability shift; and ToolMaze, which tests recovery from tool errors and corrupted results.

Across the three evaluated model sizes, the trained models improved overall on Gaia2 and ToolMaze and completed more CostBench tasks. For example, Qwen3.5-2B’s CostBench completion rate rose from 3.00% to 70.34%.

The tests use environments distinct from the training corpus, supporting the authors’ view that these demonstrations can help agents adapt beyond their training environments.

The results cover three evaluated Qwen3.5 sizes and three tests; they do not show improvement on every task or case.

The paper does not specify whether each trained model started from the exact checkpoint used for its baseline, and hidden incorrect tool results and timely responses to change remained difficult.