Source-linked AI summary
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
TL;DR
LLM agents often discard useful procedures discovered while constructing workflows, while existing executable skill libraries are usually built offline rather than from the agent’s own workflows. FlowEvo compiles successful workflows into callable skills stored in a persistent bank, reuses them directly or as workflow-generation context, and tracks their downstream utility. Across five benchmarks and ten base models, it reports the strongest overall results among the compared baselines, with evidence spanning multiple model–dataset comparisons.
Problem
LLM agents rarely retain useful procedures discovered during inference, while existing executable skill libraries typically do not grow from the agent’s own workflows.
Method
FlowEvo is a training-free framework that compiles successful workflows into callable skills, stores them persistently, and reuses them through direct execution or structured workflow-generation context.
Results
Across 5 full-split benchmarks, FlowEvo achieves the highest accuracy among 8 baselines, with gains across 10 base models and improvement over ExpeL in 49 of 50 model–dataset comparisons.
Takeaways & Limitations
Agents can improve at inference time by retaining reusable execution structure rather than relying only on transient workflow traces or fixed model parameters.
Takeaways & Limitations
The evidence is limited to settings where correctness can be checked with environment feedback or hidden tests, while fully signal-free settings remain out of scope.
Abstract
from arXiv · showhide
Large language model agents can adapt to complex tasks by constructing workflows at inference time, but procedures discovered in one episode are usually discarded after execution. Existing skill libraries provide reusable executable routines, but are typically assembled offline and do not grow from the agent's own workflows. We introduce FlowEvo, a training-free framework in which workflows and skills co-evolve at inference time. FlowEvo compiles successful workflows into callable skills, stores them in a persistent bank, and uses retrieved skills either through direct execution or as context for constructing new workflows. It also tracks each skill's downstream utility and suppresses skills that cause negative transfer. Using a shared GPT-4o-mini backbone, FlowEvo achieves the highest accuracy among 8 baselines on the full standard splits of ALFWorld, HumanEval, MBPP, GSM8K, and MATH-500. On ALFWorld, it reaches 85.6%, 26.4 points above the strongest baseline, while using roughly one third as many tokens. Across 10 base models spanning 7B to 671B parameters, FlowEvo outperforms ExpeL in 49 of 50 model-dataset comparisons. Code is available at https://github.com/DEFENSE-SEU/FlowEvo.
1 Introduction
FlowEvo addresses the gap between text-based workflow memory and offline executable skill libraries by compiling successful workflows into reusable skills that shape later solving. Its inference-time workflow–skill loop combines direct reuse with skill-conditioned generation and achieves strong benchmark performance.
- Motivation: Text memories describe prior procedures, while offline skill libraries are executable but typically fixed to a task distribution and disconnected from online problem solving.FlowEvo targets the unresolved middle ground between flexible workflows and callable, verifiable reuse.
- Method: FlowEvo compiles successful workflows into executable skill records and stores them in a persistent bank for later retrieval.The records combine callable artifacts with structured guidance.
- Method: Retrieved skills support direct execution on compatible tasks or structured context for generating new workflows when direct reuse is unsuitable.The two routes distinguish executable reuse from skill-conditioned planning.
- Results: FlowEvo attains the best accuracy on every benchmark across the full standard splits of ALFWorld, HumanEval, MBPP, GSM8K, and MATH-500 against 8 baselines.The advantage also holds across 10 base models, with controlled ablations attributing gains to the framework’s mechanisms.
- Method: Contrastive-utility lifecycle tracking suppresses skills that exhibit persistently negative transfer.This lifecycle mechanism prevents harmful accumulated skills from remaining fully active.
2 Related Work
Related work studies dynamic workflow construction, text-based workflow memory, executable tools, and training-free adaptation largely as separate directions. FlowEvo positions itself as a mechanism for persistently accumulating executable capability from the agent’s own successful workflow executions.
- Workflow generation and optimization: Workflow-optimization methods dynamically search, evolve, or route workflow structures, but the surveyed approaches do not establish persistent executable accumulation from successful executions.These methods include graph optimization, meta-agent search, MCTS, evolutionary programming, and routing systems.
- Memory for workflow agents: Text-based workflow memories guide later planning but describe prior behavior rather than providing directly executable procedures.This limitation motivates skill records that combine callable artifacts with structured guidance.
- Skills, tools, and executable reuse: Executable tools and skills are more verifiable and reusable than text-only memories, yet are typically built for fixed task distributions or outside the agent’s online loop.FlowEvo instead creates reusable skills from the agent’s own successful workflows.
- Training-free adaptation and self-evolution: Training-free adaptation improves agents through search, feedback, memory revision, or evolving control state without updating model weights.The related line lacks the concrete capability-accumulation mechanism supplied by FlowEvo.
3 Method
FlowEvo forms an inference-time loop in which tasks are solved through retrieved or generated workflows, verified traces are compiled into executable skills, and the evolving bank guides later episodes. Its skill representation, routing, admission, and lifecycle controls support callable reuse while limiting harmful transfer.
- Problem setup and loop: At episode t, FlowEvo retrieves skills, selects a route, constructs and executes a workflow when needed, verifies the outcome, and updates executable and auxiliary memory.The backbone remains fixed while the inference-time capability layer evolves.
- Skill representation: A FlowEvo skill is a structured record pairing an executable component with textual guidance, interface information, validation tests, metadata, and lifecycle state.This representation supports both callable reuse and context-based workflow generation.
- Routing and reuse: The planner routes tasks among dynamic generation, direct execution of compatible skills, and skill-conditioned generation using retrieved skills as structured context.Conditioned generation is used when retrieval helps planning but direct execution is not reliable enough.
- Workflow-to-skill compilation: Verified successful traces are compiled into callable artifacts with inferred entry points, signatures, replay tests, provenance, task-pattern tags, and routing metadata.Compilation occurs only after verifier-passed success, and candidates may be admitted for direct use, context exposure, or diagnostics.
- Lifecycle curation: The registry tracks downstream utility and failure patterns, then can restrict, repair, disable, or prune skills that accumulate negative transfer.Contrastive evaluation compares matched episodes with and without a skill and flags persistently negative deltas for suppression.
4 Experiments
FlowEvo is evaluated across five benchmarks, eight baselines, ten base models, and mechanism-focused analyses. It achieves broad accuracy gains, often with lower token costs, while reuse, adaptive generation, and curation contribute distinct benefits.
- Main results: FlowEvo attains the highest accuracy on all 5 benchmarks against 8 baselines, with the advantage holding across 10 base models.Experiments use full standard splits and shared or varied backbones to test both benchmark coverage and model dependence.
- Interactive environments: 85.6% on ALFWorld is 26.4 points above AFlow, while 9,329 tokens per episode is roughly one third of any baseline’s cost.Compiled action-sequence templates can replay directly on structurally similar tasks, where ALFWorld provides no external verifier.
- Code generation: 95.1% on HumanEval is 7.3 points above the strongest workflow-optimization baselines at 880 tokens per task, while MBPP reaches 79.6%, 5.8 points above ExpeL.Adaptive escalation solves most HumanEval tasks in a single pass and invokes more expensive strategies only after verification failure.
- Mathematical reasoning: 97.1% on GSM8K and 75.9% on MATH-500 exceed the strongest baselines by 3.2 and 2.3 points, respectively.Because these problems are effectively unique, gains come mainly from adaptive escalation and skill-conditioned generation rather than direct reuse.
- Robustness across models: FlowEvo beats ExpeL in 49 of 50 model-dataset comparisons, with mean ∆+11.0 and median +7.0 points across models spanning 7B to 671B parameters.The single below-baseline cell is Qwen3-8B on MATH-500 (−10.0), where interface-violating skills are rejected and the bank remains sparse.
- Mechanisms and ablations: Skill-to-workflow feedback produces the dominant ablation improvement, increasing performance from 38.8% to 80.6%, while curation contributes +5.0 points and suppresses harmful skills.On ALFWorld, direct reuse dominates routing; contrastive suppression prevented the harmful pick_two_obj_and_place skill from continuing to affect later episodes.
5 Discussion
FlowEvo’s accumulated skills help through direct reuse and skill-conditioned workflow generation, with benefits that are broadest for weaker backbones but bounded by feedback and bootstrapping requirements.
- Reuse mechanisms: Direct skill reuse dominates on ALFWorld, with compiled skills replayed on most episodes at a 98% success rate.This aligns with a +41.8-point skill-feedback gain over 3 seeds.
- Reuse mechanisms: Skill-conditioned generation matters when direct replay is unavailable, including code and math settings evaluated with frozen skill banks.The look_at_obj_in_light and fixed-bank MBPP analyses isolate this mode.
- Backbone scaling: 49 of 50 model–dataset comparisons favor FlowEvo over ExpeL, with a mean gain of +11.0 points and a median gain of +7.0 across the full sweep.Within the GPT-4.1 ladder, the gap reaches +53.2 points on ALFWorld for the smallest model.
- Backbone scaling: The largest gains occur where the base model is weakest, while the single trailing cell reflects insufficient surviving skills to populate the bank and support retrieval.The trailing case is Qwen3-8B on MATH-500.
- Scope and limitations: The evidence is limited to domains with checkable correctness signals, leaving free-form dialogue and long-form writing out of scope.The framework also needs enough base capability to bootstrap the loop, and delayed, partial, or adversarial feedback may pose challenges.
6 Conclusion
FlowEvo retains successful execution structure as reusable skills and evaluates its downstream effects separately from continued online adaptation through a frozen-bank protocol.
- Conclusion: FlowEvo compiles successful workflows into reusable skill records for direct execution or structured context, achieving the highest accuracy across 5 benchmarks against 8 baselines.Ablations identify skill feedback as the main contributor.
- Evaluation protocol: The fixed-bank appendix isolates reuse effects by disabling new skill compilation and other updates after the growth phase.This separates downstream effects from continued adaptation.
- Evaluation protocol: The MBPP fixed-bank analysis evaluates 71 exact-transfer cases and 44 compatible non-exact tasks across 132 episodes.Non-exact tasks are evaluated over 3 fresh task orders.
A.1 Exact-transfer analysis: executable skills as efficient direct reuse
On MBPP exact-transfer cases, executable skills improve accuracy and reduce token use relative to dynamic generation, while text memory attains slightly higher accuracy at substantially greater cost.
- Exact-transfer results: 85.9% pass rate and 204.4 average tokens favor executable reuse over dynamic generation at 81.7% and 224.2 tokens.FlowEvo selects direct execution on all 71 held-out cases.
- Interpretation: Executable skills occupy the stronger accuracy–efficiency frontier at comparable task success and additionally provide verifiability, invocability, and auditability.The comparison does not show uniform dominance at every accuracy point.
A.2 Compatible non-exact analysis: skills as structured context
The non-exact MBPP analysis tests skill-conditioned generation without direct replay and finds router effects small under fixed-bank evaluation, while broader table evidence defines the comparison settings.
- Evaluation design: Table 7 uses an episode-level priority-conflict regrouping under the same 132-episode fixed-bank budget rather than a literal task subset.This distinguishes the diagnostic from the main non-exact compatibility filter.
- Compatible non-exact evaluation: Table 6 evaluates tasks that cannot use exact replay but may benefit from related skills supplied as structured context.This isolates the context-conditioning route.
- Router analysis: All 6 router-pair comparisons differ by less than 0.6 points, with 95% confidence intervals crossing zero and McNemar p above 0.5.The tests use the full MBPP split over 3 seeds.
- Router analysis: The intended finding is that skill-as-context is robust to router choice in the no-direct-replay regime.The router effect that matters appears instead in online ALFWorld, where the bank improves look_at_obj_in_light by +28 points despite no direct execution.
B.2 Admission-gate ablation under label noise
FlowEvo’s two-stage admission and contrastive curation are designed to tolerate imperfect upstream labels while suppressing harmful skills. Ablations report small, statistically nonsignificant performance costs under removed or flipped admission labels.
- Admission and curation: A two-stage admission process combines verifier-pass compilation with contrastive checks during use to catch harmful skills that pass a corrupted gate.Downstream contrastive curation catches harmful admissions that slip through upstream label noise.
- Ablation results: 1.5–1.7 points are lost when the trace-success requirement is removed or 20% of pass/fail labels are flipped in both directions.The reported differences are not statistically significant under McNemar’s test.
- Bank growth: The bank reaches 7 workflow templates and 18 exemplars across 134 ALFWorld episodes, with the first 28 episodes covering 5 task types.The seven templates comprise six task-type templates and one blueprint entry.
- Curation analysis: Contrastive curation compares cumulative guided and holdout outcomes, but look_at_obj_in_light has no withheld episodes and therefore no defined contrastive delta.The suppression trigger is evaluated on matched guided-versus-withheld subsets.
- Task-type results: +78 points are reported for pick_heat_then_place_in_recep and +28 for look_at_obj_in_light versus ReAct across the displayed ALFWorld task types.The corresponding direct-hit rates are 83% and 0%, respectively.
C Per-Episode Algorithm
FlowEvo processes each incoming task through retrieval and routing, then either executes a compatible skill or generates a workflow with structured context. Successful workflows can be compiled, admitted, and curated into the persistent skill bank during inference.
- Per-episode loop: Algorithm 1 orders retrieval, routing, compilation, admission, and curation within one inference-time episode.The procedure is described as the workflow–skill–workflow loop.
- Retrieval and routing: The loop starts with empty skill, memory, and routing states, retrieves candidates by pattern, interface, and utility, and selects dynamic, direct-execute, or skill-conditioned routing.The task stream is processed sequentially with a fixed base model.
- Execution and conditioning: A compatible retrieved skill is called as an executable subroutine; otherwise, retrieved skills can be supplied as structured context for workflow generation.The skill-context block is empty in dynamic mode and populated in skill-conditioned mode.
- Skill compilation: Successful traces are analyzed into 3–5 abstract procedural rules emphasizing strategy, action ordering, and failure avoidance rather than exact steps.The compilation prompt excludes specific object names, locations, and numbers.
- Decision thresholds: FlowEvo’s routing, admission, and curation decisions depend on numeric thresholds, including a contrastive harm threshold of −0.1 and minimum guided and unguided sample gates of 5 and 3.A skill is suppressed only after both sample gates are met and the contrastive delta falls below −0.1.
- Skill representation: A skill record pairs an executable component with textual guidance, metadata, replay or verifier tests, and a lifecycle state.The executable component supports direct reuse, while guidance supports skill-conditioned generation.
F.2 Final skill-bank composition
After 134 ALFWorld episodes, the representative skill bank contains six task-type templates, with one template suppressed after poor guided performance. Suppressed episodes subsequently bypass the bank through dynamic routing.
- Final bank: 6 templates are compiled by the end of the 134-episode representative run, one for each task type.The pick_two_obj_and_place template is suppressed by contrastive curation.
- Suppressed template: 1/10 guided episodes succeeds for pick_two_obj_and_place, producing final utility 1/10 = 0.10 before the template is suppressed.The remaining episodes after suppression are excluded from that row’s Uses count.
- Bank accounting: The bank table counts retrieval-matched episodes as Uses, including cases routed to direct execution, skill-conditioned generation, or dynamic generation.Utility is defined as Success / Uses, with cumulative entries over the run.
G Per-Task-Type Routing Distribution
The representative ALFWorld run routes episodes among direct execution, fallback or conditioned generation, and pure dynamic generation. Direct-reuse episodes achieve high success, while contrastive curation removes a harmful task-specific template.
- Routing modes: The routing breakdown distinguishes direct_skill, direct_then_*, exemplar_guided, and pure_dynamic modes.These modes separate executable reuse, fallback, context-only guidance, and no retrieved skill.
- Direct reuse: 101/134 episodes use the direct-reuse route and achieve 99 successes, for a 98% success rate.Direct-reuse combines direct_skill with direct_then_* episodes.
- Counting convention: The reported direct-reuse count of 101 equals Direct 67 plus 34 direct_then_* episodes within Cond., not the narrower Direct column alone.Using only the Direct column would understate reuse.
- Curation outcome: A pick_two_obj_and_place template is suppressed after guided success of 1/10 versus unguided success of 1/3, with delta −0.23.The event satisfies the contrastive curation criterion in this run.
- Post-suppression routing: After suppression, 4 remaining pick_two_obj_and_place episodes use pure_dynamic routing and are excluded from the corresponding bank Uses count.They contribute to the Dyn count of 7, comprising 3 pre-suppression and 4 post-suppression episodes.