Running routine business processes with an AI agent on every case is slow and expensive because the agent calls a large language model (LLM) to decide each step. ReTrace takes the written procedure and a few recorded agent runs, including the agent's own reasoning for each action, and builds a fixed workflow. In that workflow, plain code picks the next step instead of an LLM. On 79 real-world procedures, ReTrace workflows completed 93.8% of tasks, close to the agent's 96.8%, while reducing cost per task by 24%. This makes automation practical for frequently repeated operations.
Agent traces can be compiled into workflows that are 130x faster and 24% cheaper
On 79 production procedures and two public benchmarks, the distilled workflows nearly matched the agent's accuracy.
Big Tech
Kaiyuan Hou · Yang Ping Kuo · Indranil Bhattacharya · Rohith Nama · Arushi Gupta · Carlos D. Castillo · +3 more
Columbia University · Amazon
Research Digest··3 min read
In one line
ReTrace converts SOPs and agent traces into workflows where code selects steps, cutting costs while maintaining near-agent performance.
Read it as
The authors present ReTrace, which extracts decision rules from logged agent executions that include reasoning and compiles them into a static workflow.
The authors developed ReTrace, a pipeline that takes a Standard Operating Procedure (SOP) and a few logged agent runs (traces), each including the agent's reasoning alongside its tool calls.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.