The Big Picture
Compile a few agent runs that include the agent’s reasoning into a fixed workflow and you can drop runtime planning: achieve near-agent accuracy while cutting task time from minutes to seconds and lowering per-task cost.
ON THIS PAGE
The Evidence
An agent’s step-by-step reasoning plus a handful of logged runs can be mined to recover the conditions, tool calls, and where inputs come from, then merged into a single rule document and compiled into executable workflow code. The compiled workflows run without calling a language model to plan each step, so runtime is dramatically faster and cheaper. Accuracy stays close to the original agent in many cases; with a light repair step, workflows can match or slightly exceed agent performance on some benchmarks. executable workflow code and runtime advantages contribute to throughput.
Not sure where to start?Get personalized recommendations
Data Highlights
1Time per task on a public benchmark dropped from 170 seconds to 1.3 seconds after compiling an agent into a workflow (≈130× speedup).
2Enterprise per-task cost fell by 24% when using compiled workflows instead of an agent that plans at run time.
3On the public benchmark, compiled workflows hit 75.8% task success (agent: 80.6%); after a repair step they reached 81.2%, matching or slightly exceeding the agent.
What This Means
Engineers and platform owners running high-volume automations where each run repeats the same procedure: compile once and save on compute and latency. Technical leaders deciding between keeping model-based planners in production or moving to static workflows will use these results to weigh cost and accuracy trade-offs.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The method depends on the agent providing human-readable reasoning; if the reasoning is wrong or missing, recovered conditions may be incorrect. Evaluation focused on a telecom benchmark and an internal enterprise dataset, so results may differ in other domains or for very long, branching procedures. Workflows must be rebuilt if the procedure text or available tools change, and trace selection (which branches are covered) strongly affects success on first build.
Methodology & More
Start from a standard operating procedure and a small set of recorded agent runs that include the agent’s internal reasoning alongside each tool call and its inputs. Extract the decision conditions, the tool used, and where each input came from for every step; merge these findings across traces into a compact, evidence-linked rule document. Generate executable code from that document so that, at run time, ordinary program logic picks the next step instead of calling a language model to plan. Language models are used only in the one-time construction phase. Tool Use Pattern.
Across public benchmarks and an enterprise dataset, compiled workflows achieve accuracy close to agents that plan at run time while offering huge runtime savings and lower cost. For example, a public benchmark went from 170 seconds per task to 1.3 seconds, and enterprise cost dropped 24%. The agent’s reasoning adds meaningful recovery of conditions—removing reasoning drops success by several points—so including it matters. Practical deployment requires attention to trace coverage (choose traces that exercise different branches) and to rebuilding workflows when the SOP or tools change; occasional repair after the first build can fix gaps and match agent accuracy.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
All authors have very low h-indices, no notable affiliations listed, and it's only an arXiv preprint with no citations — signals point to emerging/limited credibility.