What does 'co-evolving workflows and skills' actually mean?
Think of an AI agent as a worker who not only performs tasks but also writes down the best procedures they discover, so next time they don't have to reinvent the wheel. In technical terms, 'workflows' are the step-by-step plans an agent follows, and 'executable skills' are reusable pieces of code or actions extracted from successful runs. Co-evolution means the agent's workflow planning and its skill library improve together: the workflow uses the skills, and the skills are refined based on what the workflow needed [1][3].
For example, FlowEvo compiles successful traces into 'skill records' that include both the executable artifact and guidance on when to use it. These records are stored in a skill bank and retrieved for future tasks, creating a feedback loop: workflow-to-skill compilation, skill-to-workflow feedback, and skill curation to remove harmful skills [1]. Similarly, COSPLAY has a decision agent that retrieves skills from a learnable bank, while a separate pipeline extracts new skills from the agent's own rollouts, continually updating the bank [3].
Does it actually reduce repeated mistakes?
Yes, the evidence shows significant improvements in task success and efficiency, which directly implies fewer repeated mistakes. FlowEvo's success rate of 82.8% on ALFWorld—a benchmark for interactive household tasks—was 23.6 percentage points above the strongest baseline, meaning the agent completed far more tasks correctly. It also used less than half the average tokens per episode, indicating it didn't waste effort on failed attempts [1]. COSPLAY, tested across six game environments, achieved over 25.1% average reward improvement against four frontier LLM baselines, showing that co-evolving skills led to better long-horizon decision-making [3].
Another angle comes from a study on learning from failures: a co-evolving framework where a 'failure agent' generates hard negatives—examples that are close to success but still fail—helped the target agent sharpen its decision boundaries and improve generalization [5]. This directly addresses the problem of repeating mistakes by turning failures into structured learning signals.
Who benefits most, and what are the caveats?
The benefits are clearest in long-horizon, multi-step tasks like interactive environments (e.g., ALFWorld, games) and code/math generation, where the agent must chain many steps and can reuse successful procedures [1][3]. The frameworks are training-free (no model parameter updates), which is a practical advantage—they can improve at inference time [1]. However, the gains depend on the quality of skill curation: FlowEvo explicitly suppresses skills that cause 'negative transfer,' meaning poorly chosen skills can hurt performance [1]. Also, the studies are recent (2025-2026) and mostly on benchmarks, not real-world deployments, so real-world applicability is still emerging.
A key caveat from the process supervision study is that step-by-step feedback (process supervision) significantly outperforms outcome-only feedback for training reliable models [2]. This suggests that for co-evolution to work well, the agent needs detailed feedback on intermediate steps, not just final success/failure. So, while co-evolution is promising, its effectiveness hinges on having rich feedback signals and careful skill curation.
About These Sources
This answer is built on 5 studies (1 peer-reviewed, 4 preprints) — published from 2023 to 2026, 4 from 2024 or later — selected as the most relevant from 5 studies that passed quality screening, drawn from 28 papers retrieved from a database of over 500 million.
Sources used in this answer
FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills
FlowEvo, a training-free framework, co-evolves workflows and executable skills, achieving an 82.8% success rate on ALFWorld (23.6 points above the strongest baseline) and using less than half the tokens per episode, with ablations confirming each mechanism contributes.
Let's Verify Step by Step
Process supervision (step-by-step feedback) significantly outperforms outcome supervision for training models on the MATH dataset, with the process-supervised model solving 78% of problems, highlighting the importance of detailed feedback for multi-step reasoning.
Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Tasks
COSPLAY, a co-evolution framework with a decision agent and a skill bank agent, improves average reward by over 25.1% across six game environments compared to four frontier LLM baselines, demonstrating benefits in long-horizon tasks.
SPyCE: Skill-Policy Co-evolution for Multimodal Agents
SPyCE introduces skill-policy co-evolution for multimodal agents, where skills capture workflow-level priors for multi-step tool use, though workflow skills do not directly encode executable actions.
Co-Evolving Agents: Learning from Failures as Hard Negatives
A co-evolving agents framework with a failure agent that generates hard negatives (near-success failures) improves target agent performance and generalization by sharpening decision boundaries, showing failures can be transformed into valuable learning signals.