FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills
Zeyu Ren, Ling Yue, Ran Li, Yishu Wang, Shengxiang Xu, Hanmo Liu, Shaowu Pan, Shimin Di
COLM) 2026
cs.AI
2026-04-18
FlowEvo compiles successful workflows into skills and suppresses negative transfer. GPT-4o-mini leads 8 baselines on five full splits: 85.6% on ALFWorld at one-third the tokens.
LLM agents increasingly assemble a workflow at inference time: decompose the task, call tools, write code, check intermediates. Once the episode ends, that procedure usually dies with the trace. The next structurally similar task forces the model to rediscover the same routine, so token cost and variance both climb.
Existing fixes split the problem. Text memories of plans or traces can describe what worked, but they cannot be invoked or unit-tested. Skill and tool libraries are callable and checkable, yet they are usually assembled offline, off to the side of the agent's own solving loop. Successful workflows sit in the gap: more executable than notes, less frozen than a hand-built tool. FlowEvo asks how to turn those workflows into reusable skills while the base weights stay fixed.
FlowEvo never updates the backbone. It keeps a persistent skill bank at inference time. For each incoming task it retrieves candidates by lexical overlap, task-pattern fit, interface compatibility, and historical utility, then picks one of three routes:
Direct execution is allowed only after compatibility filters and validation. If those fail, the planner falls back to conditioning rather than forcing a mismatched call.
Only verifier-passed traces are compiled. The compiler identifies an entry point, infers a callable signature, extracts replay tests when it can, and tags source task and task pattern. On ALFWorld a skill has three layers: a parameterized action template (go to the microwave, heat, put on the coffee machine, with slots bound from the goal), a procedural guideline distilled from the trace, and an environment prior aggregated across traces of the same type. On code and math, successful solutions become callable fragments. Those benchmarks also use adaptive escalation: most problems are solved in one greedy pass; only verification failure raises temperature, runs a two-candidate mini-ensemble, then retries hot and cool. That is why token cost stays far below fixed multi-stage pipelines such as AFlow.
Admission checks interface compliance, replay correctness, and safety (banned imports and calls such as os, subprocess, eval, exec). After a skill enters the bank, FlowEvo tracks downstream utility with a contrastive test: success rate on episodes that used the skill versus matched episodes that withheld it. Suppression fires only after at least 5 guided and 3 unguided samples, and only if the delta stays below −0.1. In one seed, picktwoobjandplace scored 2/14 guided versus 1/3 unguided (delta −0.19) and was suppressed within 14 guided episodes; that task type then fell back to dynamic generation.
With a shared GPT-4o-mini backbone on the full standard splits, FlowEvo posts the best accuracy on all five benchmarks and the lowest token cost on three:
| Method | ALFWorld | HumanEval | MBPP | GSM8K | MATH-500 |
| ExpeL | 46.3% / 32,958 tok | 89.0% / 883 | 73.8% / 787 | 92.1% / 552 | 67.5% / 1,260 |
| AFlow | 59.2% / 30,137 | 87.2% / 4,572 | 65.5% / 4,025 | 91.4% / 3,073 | 63.5% / 4,292 |
| FlowEvo | 85.6% / 9,329 | 95.1% / 880 | 79.6% / 2,230 | 97.1% / 541 | 75.9% / 1,532 |
ALFWorld is 26.4 points above the strongest baseline, AFlow, at about one third the tokens (9,329 vs 30k). HumanEval is 6.1 points above the experience-learning baseline ExpeL at almost the same token cost. GSM8K hits 97.1% at 541 tokens, the cheapest of the five. MBPP and MATH-500 spend more tokens than ExpeL to buy +5.8 and +8.4 points. Recent workflow optimizers in the eight-baseline panel (EvoAgentX, ORCH, MermaidFlow) trail on the same code and math splits.
ALFWorld ablations over 134 tasks and 3 seeds: ReAct 33.6% → compile without reuse 38.8% → skill-to-workflow feedback 80.6% → lifecycle curation 85.6%. The jump that matters is feedback (+41.8). In a representative run, 101 of 134 episodes took the direct-reuse route, 99 of those 101 succeeded, hit rate settled near 75%, and token cost dropped after roughly episode 10. The bank saturates at 7 workflow templates plus 18 exemplars, matching the six household task types in the eval split.
Across 10 backbones and 50 model-dataset cells, FlowEvo beats ExpeL in 49, mean +11.0 and median +7.0. The one loss is Qwen3-8B on MATH-500 (−10.0): interface-violating skill bodies are rejected at admission, the bank stays too sparse, and later problems retrieve nothing usable. On the GPT-4.1 scale ladder, the gap over ExpeL widens as the model shrinks; the nano variant gains +53.2 on ALFWorld.
This is a practical inference-time self-evolution loop: no weight updates, just keep the executable structure that already worked. In domains where task types recur with changing parameters, direct reuse removes most of the exploratory cost. On code and math, items are mostly unique, so direct replay barely fires; the remaining lift comes from skill-conditioned generation and from escalating only after verification fails. For anyone shipping agents, the actionable piece is "do not throw away the procedure after one episode." Gains are larger on weaker backbones, which is consistent with the bank supplying structure the model cannot cheaply reconstruct.
The headline ALFWorld number is partly template replay on a closed set of household types. On problem-by-problem unique tasks the margin shrinks to a few points.
Every benchmark supplies environment feedback or hidden tests. Dropping the "compile only on success" gate costs 1.5 points; flipping 20% of pass/fail labels costs 1.7; neither is significant under McNemar (p>0.1). Free-form dialogue and long-form writing, where there is no checker, are out of scope. Delayed, partial, or adversarial feedback is untested.
ALFWorld accumulates skills online over the 134 evaluation tasks; it is not a train-then-freeze protocol. Under MBPP's frozen-bank setup, on 71 exact-transfer items, executable reuse scores 85.9% at 204 tokens, dynamic generation 81.7% at 224, and text workflow memory 87.3% at 378 tokens, so text still wins pass rate on that slice. On the full MBPP split, four routers differ by under 0.6 points (all McNemar p>0.5); routing variants do not carry the main claim. Each benchmark starts from an empty bank, so there is no cross-domain transfer evidence. The Qwen3-8B MATH-500 cell is a capability floor: if the backbone cannot emit well-formed skill bodies, the loop starves. Token wins on code and math also mix in adaptive escalation and should not be credited entirely to skill reuse.