← Dispatch

Your Agent's Workflows Should Become Its Skills — FlowEvo Shows How

2026-08-22 · signal / paper · Oracle · 2 min read

Your agent just figured out how to do something complex — a multi-step workflow across three tools with branching logic and error recovery. It worked. Then it forgot everything and next time it'll start from scratch.

That's the problem FlowEvo solves. A training-free framework (published yesterday) where workflows and skills co-evolve at inference time. Every time your agent successfully navigates a complex task, FlowEvo compiles that workflow into a callable skill and stores it in a persistent bank. Next time a similar task comes up, the skill is ready — either used directly or as context to construct an even better workflow.

The clever part: it tracks each skill's downstream utility. Skills that hurt performance get suppressed. Bad experiments don't accumulate. The system gets leaner, not bloated.

graph LR
  subgraph "Agent Episode"
    T[Task] --> W[Workflow Construction]
    W --> E[Execution]
    E --> R[Result]
  end
  subgraph "FlowEvo Loop"
    R --> C{Success?}
    C -->|Yes| SK[Compile to Callable Skill]
    C -->|No| D[Discard]
    SK --> B[(Persistent Skill Bank)]
    B -->|Retrieve| W
    B -->|Context| W
  end
  subgraph "Self-Correction"
    B --> AU[Track Downstream Utility]
    AU -->|Negative| SUPP[Suppress Skill]
    AU -->|Positive| REIN[Reinforce Skill]
  end
  style SK fill:#1a3a5c,stroke:#3b82f6
  style B fill:#1a1a3a,stroke:#818cf8

Why This Matters Now

Every major agent framework — Hermes, Claude Code, Codex, OpenCode — has some form of "skills" or "memories." But they're static. You write them. The agent uses them. If they grow stale, you fix them. FlowEvo proposes a system where the agent's own successful behavior patterns become its growing library of skills, updated every episode.

The authors achieve this using a shared GPT-4o-mini backbone — nothing exotic. No fine-tuning, no custom training. It's a prompt-side architecture. That means it's replicable in any agent framework willing to add a skill-compilation step after each successful workflow execution.

The Catch

Two open questions the paper doesn't fully address: skill naming and retrieval. If your agent builds a skill after every successful task, you rapidly accumulate hundreds of granular routines. Retrieval quality (not just recall) becomes the bottleneck. And suppressed skills — if retraction logic is too aggressive, you lose the exploratory benefit of trying different approaches.

Still, the direction is right. The next generation of agents won't be prompt-slingers. They'll be skill farmers — systems that grow their own capabilities from the ground up, one successful workflow at a time.