← Research library

Supervisor agents and the limits of agents managing agents: judgment stays with people

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

As agent programs grow from one assistant to several specialists, coordination becomes its own workload. A natural idea follows: use an agent to supervise the other agents. A supervisor can assign tasks, monitor progress, check outputs, retry failures, and decide when work is done.

Used well, this pattern can take real coordination burden off people. It also raises a question that is easy to skip. If agents supervise agents, where does human oversight actually happen, and how would anyone know if it had quietly stopped?

What supervisor agents do well

Supervisors are strongest at work that is repetitive, rule-based, and spread across many parallel tasks. We see four jobs where they earn their place.

  • Coordination: tracking many parallel tasks, noticing which have stalled, and reassigning work.
  • Consistency checks: comparing outputs from several agents for contradictions, such as two channel assets with different event dates.
  • Mechanical quality checks: verifying format, required elements, and adherence to concrete criteria.
  • Triage: sorting outputs into routine work and work that needs a person, based on defined rules.

These functions save time and catch errors that would otherwise reach human reviewers or customers. A person checking whether 10 assets all carry the same webinar date is doing work a supervisor can do faster and more consistently.

Where supervision by agents falls short

The limits are less visible than the strengths, which is why they deserve more attention. Each one tends to show up only after something has gone wrong.

Correlated errors. If the supervisor uses a similar model and similar context as the agents it checks, it may share their blind spots. A fact that fooled the writer may also fool the supervisor. Two checks that fail together give the confidence of two checks and the protection of one.

Judgment calls. Decisions about brand risk, sensitivity, fairness, and strategy still require human judgment. A supervisor can flag them but should not settle them. Our post on escalation paths covers how to define the moments an agent should stop and ask.

Accountability. When something goes wrong, "the supervisor agent approved it" is not an acceptable answer to a customer, a regulator, or leadership. A person must own the outcome, and that person needs to have seen enough to own it honestly.

Layering problems. Adding supervisors on top of supervisors increases complexity and cost. Each layer adds another place where context can be lost, and it becomes harder to reconstruct what actually happened.

How supervisors fit the hub

In the Agentic Hub pattern, the hub plans, routes, holds shared context, enforces approvals, and keeps the run log. A supervisor agent, where a team uses one, works inside that structure. It can run the consistency and mechanical checks the hub needs, and it can sort work into approval tiers.

It should not become the place where approvals are decided. Goals, guardrails, and approval rules come from people, and the hub enforces them. If the supervisor starts making calls that people configured as their own, the system has moved a judgment from a person to a model without anyone deciding to.

The same holds for self-review. A supervisor checking another agent's work is a close relative of an agent checking its own work, and it shares some of the same weaknesses. Our post on self-critique loops looks at when that kind of review helps and when it adds cost without catching much.

Design guidance

We recommend a few rules for any supervisor agent. They are aimed at the failure modes above.

  • Give supervisors different information: a supervisor with the source facts, the policies, and the original brief can check work more independently than one that sees only outputs.
  • Use supervisors to route, not to absolve: their main job should be deciding what needs human attention and making that attention efficient.
  • Keep the chain short: one layer of agent supervision is usually enough, and if you find yourself adding more, simplify the system instead.
  • Audit the supervisor: have people review a sample of work the supervisor passed, to check its judgments hold up.
  • Write down its authority: list the decisions it may make alone and the ones it must route to a person.

Different information is the most useful of these. A supervisor that can compare a draft against the approved offer record catches a wrong price. One that only reads the draft cannot know the price is wrong.

A worked example

Consider a hypothetical B2B software company launching a product update across a blog post, three emails, and a set of paid social ads. Specialist agents produce each asset, and a supervisor agent checks them before they reach reviewers.

The supervisor compares every asset against the launch brief and the shared fact record. It finds that one email names the wrong pricing tier for a new feature, and it sends that email back to the content agent. It confirms the ads meet format rules. It flags one blog paragraph that compares the product to a competitor and routes it to a person, because comparative claims are on the team's list of judgment calls.

The human reviewers now see a smaller, cleaner queue with the hard questions marked. Each month, a person samples assets the supervisor passed without escalation. In one sample, the reviewer finds a claim about data retention that the fact record did not cover, so the supervisor had nothing to check it against. The fix is an addition to the fact record and a new escalation rule, not a second supervisor.

What to measure

A supervisor is useful only if its checks hold up under human review. We suggest tracking four measures.

  • Escape rate: problems people find in work the supervisor passed, from the monthly sample.
  • Escalation precision: how often items the supervisor routes to people turn out to need a person.
  • Agreement with reviewers: how often the supervisor's pass or fail matches a reviewer's on the same work.
  • Queue reduction: how much review time the supervisor saves without raising the escape rate.

A rising escape rate is the signal to act on first. It usually means the supervisor lacks information it needs, or the work has changed in a way its rules do not cover. Supervisors also have a place on the autonomy ladder, and a supervisor whose escape rate climbs should step down a level like any other agent.

Where people sit

The division of labor we recommend is deliberate. Agents coordinate, check, and sort. People set goals, write guardrails, decide judgment calls, and own outcomes for consequential actions.

We are now studying how to make supervisor checks more independent without making them more expensive. Giving a supervisor a different model is one option, and giving it different information is another. We want to know which reduces correlated errors more, and whether the answer changes with the kind of work being checked.

How to apply this: if you use a supervisor agent, write down which decisions it may make on its own and which it must route to a person. Give it the source facts and brief, not just the outputs, and audit a sample of its approvals each month.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.