Skills mined from your agents' own runs

Every agent starts from zero and makes its own calls. Descles finds the work your agents repeat, a person reviews the procedure once, and every agent follows it after that. In our test, the same model used 76% fewer tokens with it.

What we measured

A support desk with realistic records and 16 tools. Ten refund tickets, each worked three ways by the same model (deepseek-flash, temperature 0): an agent with the tools and no guidance, the same agent with the reviewed skill, and the reviewed workflow with no model. Every model call and tool call went through the Descles edge, which counted tokens, cost and calls. A checker read the billing and support systems to decide whether each ticket was handled the team's way: the invoice refunded once, the customer answered, the ticket closed.

No playbook: 1 of 10 handled per process, 13,355 tokens and 15 tool calls per ticket, median 39 s.
With the skill: 10 of 10, 3,190 tokens (-76%) and 6 tool calls per ticket, median 5.8 s.
As a workflow: 10 of 10, no model call, 0 tokens, median 15 ms.
The skill alone, with all 16 tools still offered: 10 of 10 at 3,415 tokens, so most of the saving comes from the procedure; offering only its six tools adds the rest.

Ten tickets on one model in a demo environment, with the history mined from scripted agent runs: an illustration, not a benchmark. The run reproduces with one command, and your own agents' traces go through the same pipeline.

Why the agent without a playbook got it wrong

It was not confused. It read the ticket, the account and every invoice, then decided on its own that the charge was valid and declined 9 refunds the team's process grants. The data does not say whether to refund; that is a decision your team already made, and the agent had no way to know it. The skill carries that decision. A person still approves every refund at the edge.

How a skill is made

The edge records each run: tool names, order, decisions and outcomes, never arguments.
Descles groups similar runs and keeps the steps most of them share; occasional detours are listed, not kept.
A person binds each step's arguments and approves the draft. A draft, a rejected skill or a procedure edited after review does not run.
The reviewed procedure is written twice: a skill a model follows and can adapt, and a workflow that runs the steps with no model.
Both run through the edge, so policy and approvals still apply, and the same records measure the result.

descles-loop mine --edge https://descles.internal --admin-token-file admin-token \
  --trace-ids t1,t2,... --outcomes outcomes.jsonl --name refund-processing --out skill/
descles-loop skill review --skill skill/ --review review.yaml
descles-loop demo --live --ablation --tasks 10 --out run   # the measurement above

Questions

Skill or workflow: which should I use?

Use the workflow where the path is stable and the stakes are high: it costs no tokens and does exactly what was reviewed, but it stops at the first surprise. Use the skill where cases vary: the model follows the procedure and can adapt. Both come from the same review.

Does Descles see our runs?

Mining runs on the edge's records inside your network. The hosted control plane receives only the listed metadata the edge reports, not prompts, tool arguments or results.

Is anything learned or changed automatically?

No. Descles proposes; a person reviews and approves, and a changed procedure needs a new review. Whether a skill is adopted is decided from measured runs, not by the system.

AI agent control plane — A control plane sits beside the data plane: your agents keep talking to their model providers, and the control plane decides who they are, what they may spend, which tool calls they may make, who reviews the risky ones, and what record survives afterwards.
Claude Code governance — A coding agent with a shell is the most useful and most dangerous agent most teams run. Descles keeps the Anthropic credential off laptops and puts a policy in front of the tool calls it proposes.
Agent approvals — Autonomy is fine until the tool call is `kubectl delete pod checkout-api` in production. Approval puts one named person in that gap, and only for the calls that need it.
Documentation — endpoints, provider keys, and the full policy reference.

Try it: run the open-source edge, or create your organization and connect it. Questions: outreach@descles.com. See also terms and privacy.