ReTrace: Distilling Tool-Using Agents into Planner-Free Workflows
ReTrace turns logged agent runs into executable, planner-free workflows for repeated procedures. It trades a small amount of completion rate for lower cost and much shorter runtimes in the reported evaluations.
ReTrace extracts decision rules from an SOP and logged agent runs, then compiles them into workflows where code selects each next step. Across **79 procedures**, completion was **93.8%** versus 96.8% for the agent, at 24% lower cost. On τ²-Bench, runtime was **1.3 seconds** versus 170 seconds for the agent.
For repeated, well-defined tasks, consider turning reliable agent traces into rules and workflows instead of asking a model to plan every run. The rule document can be compiled for different orchestration platforms.
ReTrace extracts decision rules from an SOP and logged agent runs, then compiles them into workflows where code selects each next step. Across **79 procedures**, completion was **93.8%** versus 96.8% for the agent, at 24% lower cost. On τ²-Bench, runtime was **1.3 seconds** versus 170 seconds for the agent. For repeated, well-defined tasks, consider turning reliable agent traces into rules and workflows instead of asking a model to plan every run. The rule document can be compiled for different orchestration platforms. The workflow did not match the agent’s completion rate on the production procedures, and results varied by benchmark. On SOP-Bench, repair using labeled training failures raised performance to 87.8%; the paper’s results do not establish how well this approach handles changing procedures.
ReTrace turns SOPs and logged runs into executable decision rules for repeated procedures, reducing runtime and cost while retaining slightly lower completion on production procedures than the agent. Its results support compiling stable workflows, with procedure changes and benchmark variation remaining important limits.