Your AI agent is no longer you. Microsoft made it official this week: AI agents running inside companies now get their own identity — a badge of their own, with their own permissions, their own activity log, their own risk profile. They are being registered like staff, audited like staff, and restricted like staff. The company's security team has started calling them "the newest insiders."

Insiders. Not tools. Not features. Insiders — the category that includes contractors, new hires, and people with keys to things they could misuse. The classification is deliberate, and it changes the question every working professional should be asking.

For two years the question was "what can AI do?" This week it became "who approves what?" — and most people have no answer, because most people have never been told they are responsible for one.

Sixty Agents Are Already on Your Machine

Here is the detail that should get your attention. Microsoft's discovery tooling now recognises sixty distinct types of AI agents running on employee PCs — the coding assistants, the auto-drafters, the workflow runners. Claude Code. Codex. Copilot. They were found the way security teams find anything: by scanning for what is already there.

Already there. Not planned, not provisioned, not approved — there. Agents arrived the way personal phones arrived: through the side door, useful before they were governed. For months they ran as extensions of human accounts, inheriting whatever their owner could access. An agent that could read your files was indistinguishable, to the system, from you reading your files.

The badge changes that. An agent with its own identity can be given its own limits: this agent reads these folders, runs these commands, and nothing else. It can have an owner — a named human accountable for what it touches. And it can be switched off without switching you off. It is a sensible design. It is also a quiet admission: the industry now expects agents to be a permanent category of workforce, and it is building the HR infrastructure for software.

The Manager Skills Nobody Trained You For

Every insider needs a manager. That is not a metaphor — it is the emerging operating model. Agents get an owner, permissions are scoped to the task, and humans approve the actions that are hard to undo. The companies getting this right share a simple discipline: start the agent's permissions at zero, add access only where it earns its way, and keep a person between the agent and anything irreversible.

This is management. Not engineering, not prompting — management. The same skills that make someone good with a junior hire make them good with an agent: scope the work clearly, define the boundaries explicitly, review the output proportionally to the risk, and never confuse "it produced something" with "it did the job." Our Supervising AI Agents at Work course was built for exactly this discipline — it's the management layer applied to a digital workforce.

And here is the career wrinkle: nobody is going to send you a memo announcing that managing agents is now part of your job. It will just become obvious, retrospectively, that the people who did it well were more trusted, more promotable, and harder to cut. The badge your agent wears will quietly become part of your performance review.

What Owning an Agent Looks Like in Practice

Abstract governance advice is useless, so here is the concrete version. Once a week — Friday works, fifteen minutes — the owner of an agent runs the same review a decent manager runs on a report. Three checks.

Check the log. What did the agent actually do this week? Not what you asked it to do — what it did. Which files it touched, which actions it took, which external systems it called. Agents that have drifted from their brief show it in the log long before they show it in an incident. Fifteen minutes of reading beats a quarter of incident reviews.

Audit the permissions. What is it allowed to touch, and has that crept? Permissions expand the way scope expands — someone needed the agent to read one extra folder "just this once" three months ago, and the exception became the rule. The review question is simple: does the current permission set match the current task, or is it the accumulated sediment of old exceptions?

Sample the output. Pull two or three pieces of work the agent produced this week and actually check them. Not every output — a sample, chosen from the risky end: the client-facing email, the merged change, the data that fed a decision. You are not proofreading; you are calibrating how much trust the agent has earned. Trust that isn't sampled is just hope.

The Agent Lifecycle Is the Employee Lifecycle

Once you see agents as insiders, the whole management playbook transfers — including the parts nobody thinks about until they hurt. Onboarding: an agent's first week should look like a probationary hire's — narrow scope, checked output, permissions expanded only as competence is demonstrated. The teams that hand a new agent broad access on day one are making the same mistake as the manager who gives a new hire production credentials on their first morning.

Offboarding is the piece almost nobody has built. When a project ends, a vendor changes, or a tool gets replaced, the agent's access should die with it — but orphaned agents with live permissions are already accumulating in companies the way orphaned SaaS accounts did a decade ago. Every departing employee's checklist now needs a second section: not just their own access, but their agents'. The insider that keeps its badge after its owner leaves is a security incident on a timer.

And the frontier is already beyond one company's walls. Agents increasingly need to talk to other companies' agents — your procurement agent authenticating to a supplier's system, your scheduling agent negotiating with a client's. The standards for that are being written right now, and they all turn on the same question the badge answers inside a company: which agent is this, who owns it, and what is it allowed to do? Identity is becoming the control plane for the entire agent economy.

The Three Questions Checklist

If the weekly review feels like overkill for where you are, start smaller. Every agent you currently use — and you almost certainly use several, whether they wear the name "agent" or not — should have legible answers to three questions. Write them down. Genuinely write them down; the agents whose answers live in your head are the ones whose problems will too.

One: what is it allowed to touch? Which folders, systems, accounts, contacts. If the answer is "everything I can touch," that is not an answer — it is an absence of one. Two: what can it do that is hard to undo? Sent messages, merged code, deleted files, published content, money moved. Everything irreversible gets a human between it and the world. Three: who checks its work before it ships? If the answer is "nobody" or "me, when I remember," you have found the gap where the next embarrassing incident is currently forming.

One more thing worth doing before any of it: the inventory. Most people cannot list the agents they're running, because agents don't announce themselves — the autocomplete that became an auto-drafter, the plugin that started scheduling, the assistant that now summarises your inbox. Ten minutes listing every tool that acts on your behalf is usually a surprise, occasionally an alarming one. You cannot own what you haven't counted — and the inventory is the difference between governing a fleet and hoping there isn't one.

You do not need to wait for your company to issue a policy. The discipline that IT is now building into systems, you can build into your own workflow this week — and the person who already operates this way is the obvious candidate when the company formalises it.

The people who thrive in the agent era will not be the ones who use AI the most. They will be the ones who run it like a manager: named scope, deliberate permissions, proportional review. It is a new version of a very old competence — being accountable for work you delegated.

The Monday-morning version is a fifteen-minute investment: run the inventory, write the three answers for each agent you find, and schedule the first weekly review. That's it — you will have done more agent governance than most teams will do this quarter. When your employer finally issues the agent policy — and after this week, every employer will — you'll already be the person who was doing it before it was required.

We built a course for exactly this. AI Automation is a 7-day email course that takes you from casually using AI tools to actually operating them: identifying automatable work, building multi-step agent pipelines, and setting up the review points that keep delegated work safe. It pairs naturally with our recent piece on why AI agents need a manager — this week, that stopped being an opinion and became a job description. And if you want the management layer specifically — boundaries, escalation logic, accountability — our Supervising AI Agents at Work course goes deeper still.