An agent skill is a procedure. In a regulated company, procedures live under document control
Vendors now ship agents with expert-written skills and agents that learn workflows. In pharma, banking or government those are procedures, which need an owner, a version and an effective date.
Listen to this article · 5 min
AI-generated narration of the full article.
Anyone who has worked in a pharmaceutical plant knows the moment an auditor asks for the procedure that was in effect on a given day. Not the current one. The one in effect when the batch was made, with its version number, its approval signatures and the training records of the people who followed it. If it cannot be produced, the batch record has a hole in it, whatever the batch looked like.
Agent platforms are now shipping something that will face the same question, and most of them do not yet treat it that way.
Skills are procedures with a new reader
Look at what was announced this week. Google’s Gemini for government services is built around skills: reusable instructions written by domain experts, each applying the statutes, regulations and guidance that govern a task. Databricks’ Agent Memory stores, among other things, workflows that worked, distilled by a background pipeline from past sessions. Cornerstone’s new product tests whether agents follow the conduct standards of the role they occupy. Two weeks ago, skills began moving to shared libraries over MCP.
Strip away the vocabulary and each of these is a written procedure: how to perform a task, what rules apply, when to stop and ask. The only new thing is the reader. A procedure used to be read by a trained person; now it is executed by an agent.
Regulated industries already know exactly how procedures must be managed. U.S. drug manufacturing rules, for instance, require written procedures for production and process control, and require that those procedures, “including any changes,” be “drafted, reviewed, and approved by the appropriate organizational units and reviewed and approved by the quality control unit.” Deviations from them must be recorded and justified. Banks have the equivalent for credit policies and operating procedures; government agencies have it for guidance and manuals.
Our view: skills belong in the document control system’s scope
This is our position, a view rather than a fact: in a regulated company, a skill that shapes how an agent performs regulated work is a controlled document, and should be managed like one. Not a prompt in a repository that a developer edits on a Friday afternoon. Not a file the agent rewrote for itself overnight.
That does not require a new system. It requires extending the one that exists.
- Draft by the process owner, not only the developer
- Review against the regulation and test against an evaluation set
- Approve and set an effective date
- Deploy the version; every run records which one it used
- Periodic review, or a change when the rule or the model changes
- Same gates as any SOP change: Review against the regulation and test against an evaluation set · Approve and set an effective date
Owned by the business processQuality and compliance gates
Four consequences follow.
Every skill needs a business owner. The person accountable for the process, not the engineer who wrote the agent, approves what the skill says. Engineers own whether the agent follows it.
Evaluation replaces training records. For people, the evidence that a procedure can be followed is a training record. For an agent, it is an evaluation set that exercises the skill, run and passed before the version becomes effective. A skill change without a passing evaluation is an untrained procedure.
Every run must name the version it used. When an agent’s work is challenged months later, the question will be the auditor’s: which procedure was in effect? The agent’s log should answer with a skill version, the same way a batch record names an SOP revision.
Self-written procedures are drafts, never effective documents. Memory pipelines and agents that write their own skills are useful precisely because they notice patterns people miss. We argued that what agents accumulate becomes the new switching cost; in regulated work it is also unreviewed procedure. A learned workflow should land in a draft queue for the process owner, the same path we suggested for agent work that should become ordinary software, not go straight into use for everyone.
Where this view could be wrong
There are two ways. If skills turn out to be thin, interchangeable prompts with little effect on outcomes, document control would be overhead for nothing. The evidence points the other way: vendors are investing in expert-written skills because they change results. And if regulators decide that agent behavior should be validated only at the level of the whole system, per-skill control might matter less than system validation. Even then, the system’s validation would have to say which skills were in it.
Where to start
Not with every skill. Start with an inventory of the skills and learned workflows that touch regulated work, assign each an owner, and freeze the self-modification of those few until they have a review path. Then put them on the same periodic review calendar as the SOPs they mirror. The document control team will recognize the request immediately. That is the point: this is not a new kind of governance. It is an old kind with a new reader.