What does co-evolving workflows and skills actually deliver?
The core idea is that instead of starting from scratch on every task, an agent learns from its own successful problem-solving traces. FlowEvo, a training-free framework, compiles successful traces into reusable 'skill records' that pair executable code with guidance, then retrieves them for future tasks [2]. In tests across interactive environments (ALFWorld) and code/math generation (HumanEval, GSM8K), FlowEvo achieved an 82.8% success rate on ALFWorld—23.6 percentage points higher than the strongest baseline—while using less than half the average tokens per episode [2]. That means it solved more tasks with far less computational cost, directly lowering the barrier to practical use.
This co-evolution loop works because it creates a feedback cycle: workflows generate skills, and skills improve future workflows. FlowEvo's three mechanisms—workflow-to-skill compilation, skill-to-workflow feedback, and skill curation—each contribute to the overall gain, as shown by controlled ablations [2]. The skill curation part is crucial: it monitors downstream utility and suppresses skills that cause negative transfer, preventing the agent from accumulating harmful habits [2].
What's the catch? Self-evolution can make agents forget old skills.
The biggest risk is 'capability erosion'—as agents adapt to new tasks, they can progressively lose previously acquired abilities. A 2026 study found this happens across all major evolution channels: workflow, skill, model, and memory [5]. For example, in workflow evolution, optimizing for complex tasks under GPT-5.1 dropped retained simple-task performance to 41.8% [5]. That's a stark reminder that learning new things can come at the cost of old knowledge, which is a serious barrier to real-world deployment where tasks vary.
The same study proposes a solution: Capability-Preserving Evolution (CPE), a stabilization principle that constrains destructive capability drift during adaptation. With CPE, retained simple-task performance improved to 52.8% while still achieving stronger complex-task adaptation [5]. This suggests that the barrier isn't just about learning—it's about learning without forgetting, and that requires explicit safeguards.
How do we know if it's actually working? Evaluation is still a hurdle.
Even with co-evolution, current agents are far from reliable in real-world workflows. A live benchmark, Claw-Eval-Live, tested 13 frontier models on 105 tasks spanning business services and local workspace repair; the leading model passed only 66.7% of tasks, and no model reached 70% [1]. This shows that while co-evolution can boost performance, it doesn't yet make agents dependable enough for autonomous use in complex, multi-system workflows.
The benchmark also highlights that leaderboard rank alone is insufficient—models with similar pass rates can diverge in overall completion, and task-level discrimination concentrates in a middle band of tasks [1]. This means we need evaluation that goes beyond final answers, using execution traces and audit logs to verify actual task completion [1]. Similarly, in autonomous research, a 2026 study found that executable workflows alone don't produce research judgment; systems need to capture trial-and-error experiences and convert them into behavior changes [3]. Their framework, Sibyl-AutoResearch, showed that such conversion is possible, with eight high-confidence events recovered in a retrospective audit [3].
When does co-evolution help most, and what's needed next?
Co-evolution seems most effective in domains where tasks are repetitive enough to benefit from reusable skills, like interactive environments and code generation [2]. But in highly fragmented fields like architecture, engineering, and construction (AEC), the barrier isn't just about skills—it's about infrastructure. A 2026 paper argues that AEC needs 'agent-native infrastructure' that couples semantic knowledge, tool interaction, and policy-constrained workflow execution to turn fragmented project knowledge into executable assets [4]. This suggests that co-evolution alone can't overcome systemic fragmentation; it needs a supportive environment.
Looking ahead, the evidence points to three priorities: (1) building safeguards against capability erosion, as CPE demonstrates [5]; (2) developing evaluation methods that verify real-world execution, not just final answers [1]; and (3) creating infrastructure that supports knowledge reuse and governance [4]. If these are addressed, co-evolving workflows and skills could indeed lower the barrier to advanced AI agents—but it's a gradual process, not an overnight fix.
About These Sources
This answer is built on 5 studies (4 peer-reviewed, 1 preprint) — published in 2026, 5 from 2024 or later — selected as the most relevant from 5 studies that passed quality screening, drawn from 45 papers retrieved from a database of over 500 million.
Sources used in this answer
Claw-Eval-Live: A Live Agent Benchmark for Evolving Real-World Workflows
Claw-Eval-Live, a live benchmark with 105 tasks, found that the leading model passed only 66.7% of tasks and no model reached 70%, indicating that reliable workflow automation remains unsolved and that evaluation must verify execution traces, not just final answers.
FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills
FlowEvo, a training-free framework that compiles successful traces into reusable skill records, achieved an 82.8% success rate on ALFWorld (23.6 percentage points above the strongest baseline) while using less than half the average tokens per episode, demonstrating that co-evolving workflows and skills can improve accuracy and efficiency.
Sibyl-AutoResearch: Autonomous Research Needs Self-Evolving Trial-and-Error Harnesses, Not Paper Generators
Sibyl-AutoResearch, a self-evolving research framework, showed that executable workflows alone don't produce research judgment; it introduced trial-and-error harnesses that convert trial signals into behavior changes, with a retrospective audit identifying eight high-confidence conversion events.
Toward agent-native infrastructure in AEC: From fragmented knowledge to executable workflows
In the AEC sector, a 2026 paper argues that the next bottleneck is not digital tools but the lack of agent-native infrastructure that couples semantic knowledge, tool interaction, and policy-constrained workflow execution to turn fragmented project knowledge into executable assets.
Do Self-Evolving Agents Forget? Capability Degradation and Preservation in Lifelong LLM Agent Adaptation
A study on self-evolving agents found that adaptation to new tasks can progressively degrade previously acquired capabilities across all evolution channels; with Capability-Preserving Evolution (CPE), retained simple-task performance improved from 41.8% to 52.8% while preserving adaptation performance.