The Agent You Approved Is Not the Agent Running

On October 15, Microsoft announced a set of Copilot Studio features in its Microsoft Copilot Blog post on persistent, proactive agents: agent accounts with their own credentials, business process memory, and triggers that start an agent on an external signal or a set cadence, no prompt required. The features are rolling out through the Frontier program.
Read the memory description closely. Business process memory lets an agent carry forward preferences and process context into each new cycle, and Microsoft puts it plainly: any idiosyncrasy of the business process that one person raises to the agent becomes a shared resource for everyone else. One person’s aside, offered in one conversation, becomes operating context for the whole team in the next cycle.
Microsoft also describes a feature planned for the coming months called agent refinements. The agent’s own activity will be used to recommend improvements to its maker, who evaluates, tests, and implements them. The recommendations can convert repeated steps into workflows, capture reusable approaches as skills, flag ambiguity in the instructions, or suggest a different model for the task.
Put those three pieces together and a plain fact falls out. The agent that went through your approval process is a snapshot. The agent running next month has different memory, possibly different instructions and a different model, new ways to start work on its own, and a wider set of people whose input shapes how it behaves. None of that required a new approval. It just required time.
Approval processes assume a stable thing being approved. Software teams learned to handle this decades ago. Nobody approves an application once and then lets it change freely in production. A change to code, configuration, data sources, or the user population goes through a release: a version, a diff, a test, a rollback plan. Agents need the same treatment, and most shops are not giving it to them. They approve the agent at launch and then let three kinds of change accumulate without a record.
The platform changes underneath it. Agent accounts, shared memory, and always-on triggers did not exist in this form when many Copilot Studio agents were first approved. A platform release can hand an existing agent a new capability, a new way to be invoked, or a new place its output lands. Your approval said nothing about that capability because it was not there to evaluate. When a vendor ships a capability change to a platform your agents run on, that is a change to your agents, whether your team touched anything or not. It belongs in the same review queue as a change your own makers propose.
Memory writes are changes. In a conventional system, changing behavior requires editing code or configuration, and both leave a diff. An agent with shared process memory changes behavior when a user teaches it something. That teaching is an unreviewed edit, made by whoever happened to talk to the agent, with no diff, no test, and no expiry unless someone builds one. Microsoft’s design does include a way to view and delete saved memories, which is the right instinct. Treat it as a change log, not a settings page. Know who can add to shared memory. Review what has accumulated. Expire entries that were true for one cycle and wrong for the next. A wrong fact in shared memory does not sit in one conversation. It propagates to every cycle that follows, attributed to the process itself.
I made a related argument in AgenticOps Is Here. Drift Is Too.: granted autonomy drifts from observed behavior over time. Memory is one of the mechanisms that produces the drift, and unlike model drift, this one is written by your own users, one helpful correction at a time.
Accepting a refinement is a release. The refinement loop Microsoft describes is sensible. An agent that notices repeated steps and proposes turning them into a workflow is doing useful work. But the moment a maker accepts a suggestion that changes instructions, adds a skill, or swaps the model, the agent’s behavior changes for every user and every future run. That click is a production release. It deserves the same handling as any other release: what changed, what test ran against the change, what the previous version was, and how to get back to it if the new behavior is worse. A suggestion generated from the agent’s own activity is still a change proposal from an interested party. Test it like one.
So version the agent as a unit. The version is not just the prompt. It is the instructions, the model, the tools and their permissions, the data sources, the triggers, the memory policy, and the population allowed to use it and teach it. Change any one of those and you have a new version. That is the discipline I argued for in Git Is the Control Plane, applied to the change record itself: if the change is not recorded in a form you can diff and roll back, it did not go through a process. It just happened.
Set re-approval triggers in advance, while the agent is still small. A new data source, a new trigger that starts work without a person asking, an expansion from one team to the whole department, a model swap, a new tool with write access. Each one reopens the approval, not as bureaucracy, but because each one changes what you actually approved. I covered the boundary side of this in Always-On Agents Need Declared Boundaries: a boundary declared at launch only holds if something watches for the changes that cross it.
None of this argues against persistent, proactive agents. A budget agent that carries unresolved questions from one review cycle to the next, which is Microsoft’s own worked example, is genuinely useful. The point is narrower. Persistence and self-improvement mean the thing you run diverges from the thing you approved, by design, every week it operates. Your governance has to run at the same cadence as the change, or it is governing a system that no longer exists.
Run one test this week. Pick an agent in production and ask its owner to name the version running today: instructions, model, tools, triggers, memory contents, user population. If the owner cannot produce that in one page, you do not have a change problem yet. You have a simpler one. You cannot say what you approved, because nobody wrote it down.
Thinx-Tank