Business process automation runs the same Standard Operating Procedure again and again on new inputs: triage a support case, check a form, escalate an order. LLM agents can follow an SOP written in plain natural language — but they call a model to choose every step of every run, so latency and cost scale with volume and the context window grows with each step.
ReTrace (Columbia University + Amazon) flips that. It takes the SOP plus a handful of logged agent runs and compiles them into a reusable workflow where code, not an LLM, selects the next step. The insight is that an agent’s rationale — the explanation it writes alongside each tool call — usually states why it acted, while the tool calls alone do not. ReTrace extracts the condition behind each step, merges conditions across runs into a compact rule document, and generates an executable workflow from it. Only this last stage is platform-specific, so one rule document can be compiled for different orchestration platforms.
Production numbers. On 79 procedures from a production platform, ReTrace workflows reach 93.8% completion versus 96.8% for the original agent, at 24% lower cost per task. On tau2-bench, task success is 75.8% (agent: 80.6%), but runtime collapses from 170 s to 1.3 s — the compiled workflow does the same work without a planner call per step.
Repair closes the gap. Three rounds of repair using labelled training failures lift k=2 workflows from 66.7% to 81.2%. On eight SOP-Bench procedures, ReTrace without repair trails few-shot generation (80.6% vs 83.6%), but after repair it reaches 87.8% at 0.40 s per task, versus 85.5% in 46.6 s for the agent.
Why it matters beyond the benchmark. The agent’s own rationale is doing double duty: it is both the explanation shown to operators and the raw material for compilation. Memory-augmented approaches throw that reasoning away and keep only hints; ReTrace keeps the reasoning and turns it into a rule document — a small, reviewable artifact that a human can approve before it runs in production. That is a meaningfully different relationship between an agent and a business process than “call the model again and hope”.
The trade-off is explicit: you give up a few points of completion for a workflow that is cheaper, ~100x faster, and auditable as code. For high-volume, repeated SOPs, that is an easy bargain — and a template for “compile the agent once, run it many times”.
Paper: arXiv:2611.14492