ReTrace: Distilling Tool-Using Agents into Planner-Free Workflows
On 79 production procedures, ReTrace cut cost per task by 24%, with completion falling from 96.8% to 93.8%.
ReTrace extracts decision conditions from agents’ explanations of their tool calls, merges them into rules, and generates workflows whose next steps are selected by code. On τ²-Bench, those workflows ran in 1.3 seconds versus 170 seconds for the agent, with task success of 75.8% versus 80.6%. Repairs based on labeled training failures improved results: on eight SOP-Bench procedures, repaired workflows reached 87.8% success at 0.40 seconds per task, compared with the agent’s 85.5% in 46.6 seconds.
ArXiv · AI/CL/LG's note
ReTrace extracts decision conditions from agents’ explanations of their tool calls, merges them into rules, and generates workflows whose next steps are selected by code. On τ²-Bench, those workflows ran in 1.3 seconds versus 170 seconds for the agent, with task success of 75.8% versus 80.6%. Repairs based on labeled training failures improved results: on eight SOP-Bench procedures, repaired workflows reached 87.8% success at 0.40 seconds per task, compared with the agent’s 85.5% in 46.6 seconds.
ArXiv · AI/CL/LG's note
score 5