Analyses & Essays

When should an AI agent get a team of its own?

Suppose you’re choosing a conference venue for 180 people. The lowest bid includes a ballroom that seats 200. It looks like an easy choice until the program director asks, “How many seats are on the other side of the divider?”

You need two simultaneous workshops for 90 people each. The question changes what counts as evidence. Someone has to find the divided-room floor plan, check the seating arrangement, and go back to the bid to see which rooms the price actually covers. You could have given an AI agent a checklist for comparing venues. Would you have thought to put that sequence on it?

This is where I’d consider giving a sub-agent a team of its own. An agent can inspect what it finds and choose what to do next. In a nested system, that choice can include putting another agent in charge of a smaller question, with its own helpers and a clear answer due back. The organizer sets the requirements once. The agents can work out which follow-up questions each bid deserves.

Follow one bid down and one answer back

Imagine three hotels submit proposals. The lead agent gives each bid to a separate coordinator. Two hotels offer separately named workshop rooms with capacities and prices in their bids. Those coordinators have enough to finish. At the third hotel, one ambiguous line about a ballroom divider sets off a different chain of work.

Who decides what on the third hotel’s bid?
  1. Comparison lead. Sets the common requirement and asks for a shortlist the conference organizer can use.
  2. Hotel coordinator. Spots the divider and asks a room-layout agent to establish how many people each half can seat for a workshop.
  3. Room-layout agent. Assigns one helper to read the seating chart and another to identify the rooms on the floor plan. Checks that both used the same layout and room names.
  4. Source-checking agents. Return the marked-up chart and plan. Together the documents show 110 seats on one side and 70 on the other.

Now the hotel coordinator has a precise follow-up. It sends its pricing agent back to the quote to check whether the organizer also gets the 100-seat garden room shown on the plan. The room is available only as an unpriced extra. The coordinator sends that finding, the quote line, and the two source pages up to the lead. The shortlist can now say, “Two bids meet the brief at the quoted price. Ask the third hotel to price the garden room before comparing it.” The organizer can make that call.

What did the extra layer buy? The room-layout agent could settle a local question that took two documents to answer. Its finding changed the pricing agent’s assignment. Neither of the other hotel teams had to do that work, and the lead could compare three bids without reading three sets of floor plans.

Sometimes the right move is to stop several agents

Consider a different hypothetical. A bicycle-light maker changes how riders switch modes, from a double tap to a long press. Before the next print run, a release agent asks one coordinator per language edition to find instructions that still teach the old gesture. In the French edition, the text is correct. The illustration still shows a finger tapping twice.

The French coordinator asks an illustration agent to trace the drawing. One helper finds the image identifier in the page layout. Another checks which editions use that same asset. All six do. Once that finding reaches the lead, it can stop the other coordinators from commissioning local redraws. It requests one correction to the master illustration, then sends the revised proofs back for a caption check. The editor who signs off can see which proofs passed and which still need attention.

That discovery changes who owns the work. If every coordinator quietly fixed its own copy, six editions might look fine on press day while the old drawing sat in the master library, ready for the next print run.

Where would a second layer earn its keep?

Pull up the last piece of work your team sent back for another pass. Find the comment that reopened it. “Please check the numbers” tells you little. “Does this price include the room we need at two o’clock?” gives you a place to start.

  1. What could the reviewer decide independently? Give the agent a question it can close, the sources it can use, and a way to say the evidence runs out. In the venue example, the agent can establish room capacity. Only the organizer can relax the 90-seat requirement.
  2. What would make it call in help? Name the finding that would change the next step. A divided room triggers a layout check. An old drawing triggers a search for other editions that use it. If every case follows the same three steps, a checklist may already do the job.
  3. What has to come back? Ask for the answer, the source that supports it, and any open question someone must resolve. “Venue passes” loses the condition on the garden room. “Passes if the extra room is included” gives the buyer something to act on.

Extra agents cost money and can duplicate one another’s mistakes. Anthropic’s research team, for example, saw agents assigned a broad research question repeat the same searches. Giving each agent a specific assignment helped. Read the team’s account.

I’d test the design on a handful of finished jobs, with the final decisions hidden. Give a single agent and the nested version the same source materials. Have the usual reviewer mark missed exceptions and time the corrections. Count the model bill and the reviewer’s time per accepted result. If delegation catches a consequential miss, inspect the handoff that caught it. That’s the part worth testing on the next batch.

If you have a recurring review like this, bring one anonymized example and the comment that sent it back. I help teams turn that pair into a pilot with a clear question, a baseline, and a result they can judge for themselves. Talk with me about your workflow.