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

Agents That Learn Their Own Shortcuts

FlowEvo tackles a fundamental problem with LLM agents: they solve problems from scratch every time, even when they've already figured out good solutions. The framework lets agents extract reusable skills from successful workflows during inference itself, then automatically apply or build on those skills in future tasks—without any offline training.

Most existing approaches split the world into two categories. Workflow-based agents (like chain-of-thought or tree-search methods) construct task-specific procedures on the fly but discard them after use. Skill libraries offer reusable routines—but someone has to manually curate them offline. FlowEvo merges both: the agent's own successful problem-solving episodes become the seed for a dynamically growing skill bank.

How It Works

When a workflow solves a task, FlowEvo compiles it into a callable skill and stores it. The next time the agent faces a new problem, it can retrieve related skills from the bank in two ways. It either executes them directly (like calling a function) or feeds them as examples to GPT-4o-mini, which uses them as context when constructing a new workflow. This dual use—both as executable routines and as in-context demonstrations—is key.

Critically, the framework suppresses skills that hurt performance. Each skill tracks its downstream utility: if using it leads to worse outcomes on downstream tasks, the agent learns to avoid it. This prevents negative transfer, a common failure mode when reusing knowledge blindly.

The whole process runs at inference time with a single GPT-4o-mini model. No separate training loop, no offline skill curation—just the agent progressively refining its own toolkit.

Results That Matter

Across five diverse benchmarks (ALFWorld, HumanEval, MBPP, GSM8K, MATH-500), FlowEvo achieves the highest accuracy among 8 baselines when using a shared GPT-4o-mini backbone. On ALFWorld—a simulated household task environment—it reaches 85.6%, a 26.4 percentage point lead over the strongest competitor. Even more striking: it uses roughly one-third as many tokens, meaning faster and cheaper inference.

The gains hold across model scales. Testing on 10 different base models (7B to 671B parameters), FlowEvo outperforms ExpeL, a strong baseline that also evolves skills, in 49 out of 50 model-dataset comparisons.

Why This Matters

This work dissolves a false choice in agent design. Rather than choosing between dynamic, problem-specific workflows (flexible but wasteful) and static, curated skills (efficient but rigid), FlowEvo grows skills organically from the agent's own successes. The agent becomes less like a worker following a fixed playbook and more like a practitioner who internalizes shortcuts from past experience—and knows when not to apply them.

The token efficiency gain is significant for practical deployment: fewer API calls mean lower latency and cost. That the approach is training-free is also important—no need to collect new datasets or fine-tune models, just run it.