FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills

Most agent frameworks treat each episode as disposable: the model reasons, acts, succeeds or fails, and the useful structure evaporates. FlowEvo closes that loop. When an episode succeeds and passes verification, the system compiles the workflow into a callable skill record—an executable artifact with an interface, replay tests, and metadata—and stores it in a persistent bank. Later tasks can either execute that skill directly or receive it as structured context that shapes how a new workflow is built. The base model's weights never change.

The distinction between those two reuse modes matters more than it might seem. On ALFWorld, a text-based household environment with six recurring task types, direct execution dominates: compiled templates replay on the majority of episodes at a 98% success rate in the representative run, and average token cost drops to roughly a third of any baseline's figure. That 85.6% overall accuracy sits 26.4 points above the strongest baseline. On math benchmarks like MATH-500 and GSM8K, where each problem is effectively unique, direct replay contributes almost nothing—the gain comes from skill-conditioned generation, where retrieved skills supply accumulated solution patterns as structured context without being replayed verbatim.

The system's curation mechanism is worth examining closely. FlowEvo tracks each skill's downstream utility through contrastive evaluation: it periodically compares success rates on episodes where a skill was applied against matched episodes where it was withheld. When the delta is persistently negative, the skill is suppressed. In one seed of the three-seed ALFWorld campaign, the pick_two_obj_and_place skill was flagged—guided success was 2 out of 14 episodes versus 1 out of 3 unguided, a delta below the suppression threshold—and removed from active retrieval. Without that mechanism, the bank would amplify its own worst entries.

The robustness results are notable. Flipping 20% of pass/fail labels in both directions costs only around 1.5 to 1.7 accuracy points and is not statistically significant under McNemar's test. The two-stage design—verifier pass at compile time, contrastive checking during use—absorbs upstream label noise without requiring clean supervision. Across ten base models spanning 7B to 671B parameters, FlowEvo beats ExpeL in 49 of 50 model–dataset comparisons, and the gain over ExpeL widens as model size decreases on ALFWorld, reaching its largest margin on the smallest model tested. The capability layer substitutes structure where the backbone lacks it and contributes less where the backbone can already reconstruct that structure independently.

The one failure case is instructive: Qwen3-8B on MATH-500, where too few candidate skills survive admission to populate the bank usefully. That floor is real. So is the scope limitation: the framework requires verifiable correctness signals, and fully signal-free settings like free-form writing remain out of scope. Bank maintenance under long-running use—merging, retiring, and revising skills without accumulating retrieval interference—is identified as an open problem rather than a solved one. Those are honest constraints on a system that otherwise demonstrates a clean mechanism for persistent inference-time capability accumulation.