When Adding a Supervisor Agent Makes the System Worse

The diagram looks more capable after the change. Instead of one AI research assistant, it shows several specialists and a supervisor assigning work and combining their findings. It resembles an organized team, which makes the design easy to like.

Then the same research task takes twice as long and loses an important citation.

That is an illustrative result, not a measurement from a client. It raises a concrete architectural question: what did the additional coordination contribute to the answer the user received?

Give the specialists a job worth separating

A supervisor agent delegates work and may assess or combine the results. Separation can help when specialists need different tools, permissions or bodies of information. It is less convincing when every agent has the same capabilities and the supervisor merely chooses a sequence already known in advance.

There are several legitimate arrangements to compare. One agent can use defined tools. A fixed workflow can pass through search, checking and drafting steps. A supervisor can choose the next contribution as findings arrive. OpenAI's orchestration guidance distinguishes model-led orchestration from control expressed in code; a fixed sequence can still use AI inside its steps. [1]

I would choose between those arrangements from the work, not from how closely the diagram resembles a department.

A missing citation can reveal the wrong repair

In our research example, the assistant is comparing two document-processing services for a confidential workflow. A specialist finds a provider document saying that a retention exception is available only under a particular agreement. It returns the finding, the source and the qualification. The supervisor's final comparison says simply that the service offers the exception, omitting the citation and its condition.

The search succeeded and the source reached the supervisor, but the loss happened when the findings were combined.

Asking the specialist to search harder would target the wrong stage. Adding another critic might produce more discussion without changing the handoff that stripped out the condition. I would first preserve the finding, source and qualification as one contribution, then check whether the final comparison retains them.

The finished answer is the product the user relies on. Strong results from individual workers don't establish that the combined system preserved their evidence. Evaluations should therefore include the whole task as well as the components that help explain a failure. [2]

When a supervisor genuinely changes the route

Now change the research situation. One source reveals that the proposed deployment depends on whether customer documents may leave a specific environment. That finding creates a question the initial research plan did not settle. A coordinator could ask a specialist with access to the relevant internal requirements to examine it, then request a different technical comparison based on the answer.

The route now depends on an intermediate finding. Dynamic coordination has a recognizable job: choosing what to investigate next rather than repeatedly choosing the same predefined step.

That gives dynamic coordination a purpose to test, not proof that it will outperform a fixed workflow. Where practical, keeping the workers and task constant while changing the coordination helps distinguish that benefit from simply granting the system more attempts or a stronger model.

Time and cost belong beside answer quality. Taking twice as long might be worthwhile for a substantially better-supported decision. Taking twice as long to deliver a less faithful version of the same evidence is a reason to reconsider the design.

More iterations need a remaining problem

A missing source or unresolved contradiction can justify another pass. Rephrasing an already supported answer may not. The workflow needs a way to distinguish useful additional work from a conversation that continues because another critique is possible.

A limit on time, cost or iterations gives that comparison a practical boundary. If a material uncertainty remains when the limit is reached, the result should preserve it rather than manufacture agreement to end neatly.

For the original twice-as-long, source-losing example, I would restore the simpler arrangement that meets the task while repairing the synthesis problem. Supervision can return where it addresses a need the baseline handles poorly. It may belong in one part of the workflow rather than everywhere.

Removing the supervisor is not a retreat from AI. When fewer handoffs deliver the required answer more reliably, the team has learned something useful about the work. The architecture should be allowed to become smaller as that understanding improves.