Stop Losing the Workflow Your Agent Just Discovered

The short version

When an agent finally solves a messy task, the useful thing is often not the final answer.

It is the workflow it discovered on the way there.

FlowEvo gives that idea a concrete shape. The current arXiv version describes a training-free framework where successful workflows are compiled into callable skills, stored in a persistent bank, and retrieved for later tasks either as executable routines or as planning context. The paper’s useful claim is not just that agents should “remember” more. It argues for skill records with a route, evidence, and lifecycle control.

That is the part worth stealing.

Do not let a working agent trace evaporate into chat history. Turn the verified part into a small skill that future runs can call, inspect, or ignore.

What changed

Most agent memory is still text. It can remind the model what happened, but it is hard to verify and easy to overtrust.

FlowEvo treats the durable object differently. A stored skill is not just a note saying “last time, do this.” It has an executable body, a callable interface, replay tests or verifier evidence, metadata about where it came from, and a lifecycle state. Later, the agent can either execute the skill directly or use its structured description as planning context.

That distinction matters. A direct skill should be held to a higher bar than a hint. If checks are unclear, keep it as context or leave it in a shadow state instead of trusting it as a tool.

The workflow to steal

Use this for recurring coding-agent, browser-agent, ops-agent, or assistant tasks.

1. Run the agent with a real completion gate: tests, build, browser check, verifier, human review, or policy checklist.
2. When the run succeeds, isolate the reusable trace segment. Ignore the motivational commentary and keep the operational move.
3. Define a tiny interface: trigger condition, inputs, outputs, allowed tools, preconditions, and fallback path.
4. Add one replay check that proves the skill still works outside the original chat.
5. Admit it as draft-only until a human owner reviews provenance and safety.
6. Track whether future runs improve. If the skill causes wrong shortcuts, downgrade or delete it.

FlowEvo calls this the negative-transfer problem. A stored skill can help future tasks, but it can also bias the agent toward the wrong shortcut. The interesting system move is the governance layer: keep measuring downstream utility and suppress skills that hurt later runs.

A prompt for the next successful run

From this successful agent run, extract one reusable skill candidate.

Include:
- trigger conditions
- inputs and outputs
- the callable artifact or exact workflow
- replay test
- safety preconditions
- provenance from this trace
- when to use it only as context
- when to retire it for negative transfer

Reject the skill if success depended on hidden state, luck, stale data,
private credentials, or an unverifiable assumption.

The useful habit is the rejection step. A skill bank is only better than memory if it refuses weak memories.

What this replaces

For a small team, start with a folder of draft skills, one replay command, and a review rule: no successful trace becomes active just because the model liked it. Keep “context-only” as a real state. Some traces are useful as hints but too brittle to run directly.

Where it fails

The failure mode is obvious: you automate your bad habits.

If a trace passed because the cache was warm, the hidden fixture was forgiving, or the model guessed the right string once, preserving that as a skill makes the next run worse with more confidence.

The practical lesson from FlowEvo is sharp: agent improvement does not have to mean model training.

Sometimes it means preserving the one verified workflow your agent already earned, then being willing to remove it when it stops helping.

Sources

• FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills
• DEFENSE-SEU/FlowEvo public repository
• Private Huecki radar artifact: content-research/radar/2026-08-23.md