Should Agents Have Performance Reviews? A Debate
An agent with an employee ID
SK Telecom just gave its AI agents employee ID numbers. In its first AI governance report, the Korean carrier says it formally recognizes agents as "a new form of labor." Each agent has a role, a supervisor, an operating status and tracked performance.
The company goes further. Agents judged less necessary based on evaluation results will be suspended or retired. It is "reviewing ways to share agents' performance with our employees," and over the longer term will consider reflecting agent performance in organizational KPIs.
In plain terms, agents are getting something that looks a lot like a performance review. That raises a question most organizations haven't asked out loud yet: should they? This post lays out both sides. Then we want to hear from you.
The case for reviewing agents
What isn't measured isn't managed. Agents now do real work inside real workflows. If a person's output is reviewed but an agent's isn't, the agent becomes an unexamined cost center. SK Telecom already tracks token usage and work records for each agent. A review is just the habit of looking at that data on a schedule and acting on it.
Retirement needs a standard. Agents multiply fast. Without a regular evaluation, nobody decides which ones are still earning their keep, and the organization ends up with dozens of half-used agents nobody owns. A review creates a clear moment to keep, fix or switch one off.
Visibility builds trust. Sharing agent performance with employees could do the opposite of what critics fear. People who can see what an agent actually handles, and how often it gets things wrong, are better placed to rely on it, check it, or push back on it.
It forces the accountability question. Giving an agent a named supervisor, as SK Telecom does, means someone has to answer for its results. A review makes that ownership concrete instead of theoretical.
The case against
An agent can't underperform on its own. An agent's output depends on the instructions it was given, the data it can reach, and the workflow it sits in. When results are poor, the cause is usually upstream: a vague handoff, a bad prompt, a missing review step. Reviewing the agent points at the wrong thing.
It invites the wrong comparison. Once agent results sit next to human results in the same KPIs, people will be measured against the software beside them. That's a recipe for anxiety and gamed numbers, not better work. Agents and people do different jobs in the same workflow, and they need different measures.
"Performance review" is the wrong frame. Reviews exist for people: growth, feedback, pay, careers. An agent has none of those. What it needs is closer to monitoring and quality control, with a clear owner. Borrowing HR language blurs who is actually responsible.
It can hide human accountability. If an agent gets a low score, it's easy to retire the agent and move on. The person who designed, deployed and supervised it never has to explain what went wrong.
The real question: who owns the work?
Both sides agree on one thing. When an agent does work, someone has to own the result.
SK Telecom's answer is a named human supervisor, such as a team leader or executive, for agents with more impact or autonomy. Human review and approval tighten as autonomy grows. That's a sound starting point, and it's the part most organizations skip.
The harder design question is what that supervisor is accountable for. Is it the agent's output, like a manager answering for a direct report? Or the workflow the agent sits in, including the handoffs, checks and data around it? The answer decides what you measure. Measure the agent alone, and you'll keep swapping agents. Measure the workflow, and you'll find out why the agent struggled in the first place.
Maybe the review isn't for the agent at all. Maybe it's a review of the workflow, with the agent as one participant.
Where do you land?
We'd like to hear from practitioners who are already running agents in real workflows:
Should agent performance ever sit in the same KPIs as human performance?
When an agent underperforms, who should answer for it: its supervisor, the person who built it, or the owner of the workflow?
Would you want to see the performance of the agents you work alongside?
Is "performance review" the right frame, or does it need a different name and a different process?
Share your take in the comments or on LinkedIn. The strongest responses may show up in a future post.
Sources
SK Telecom issues first AI governance report, Seoul Economic Daily, October 26, 2026 (English translation of the Korean original)


Comments