Opening

An agency built a virtual team to produce market reports. Five agents. A senior researcher, a data analyst, a writer, a reviewer, and an editor in chief.

Look at the configuration. Only the researcher has a tool, web search.

All the others have none. And the flow is always the same sequence.

The trace shows what happens. The researcher runs the searches, and writes a summary. The analyst rewrites the summary in other words. The writer turns it into a report. The reviewer changes a few sentences. And the editor replies: approved. Five large calls, each rereading the previous one's text. A single agent, with the search tool, delivers the same report. In this lesson, you'll learn to tell decoration from real division of work.

The diagnosis

Here's the shape. The only agent with a tool does the real work.

The others read the whole text, and return a similar version.

And the result costs five times as much, and takes five times as long.

How do you spot it? The agents have human job titles, like researcher, analyst or manager, instead of technical responsibilities.

Only one or two agents have tools. The others only read, and rewrite.

The flow is always the same sequence, with no decision about who works.

Each agent receives the previous one's entire text, and returns a similar text.

And removing an agent doesn't change quality in the eval. While the cost is several times that of one call.

The ablation test

The honest way to see this smell is to drop an agent, and measure. Say the agency's eval has forty reports, and a rubric. Claims with sources, numbers that match the page, the agreed structure.

You turn off the reviewer and the editor in chief. The score doesn't move.

You turn off the analyst. The writer works straight from the researcher's summary, and the score stays the same.

One agent that researches and writes is left. The other four were job titles.

And if the score drops when you remove someone, that split is paying for something. Then you name the technical reason, and keep only that.

A challenge

Here's one for you. You turn off the reviewer and the editor in chief, on an eval of forty reports. The rubric score doesn't change. What's the most useful diagnosis?

Those two roles are decorative. They reread and rewrite, without doing any new work. Collapsing them is the next step. Not reinforcing the job title.

A bigger model on the reviewer, with no criterion of its own, only makes the rewrite more expensive. And adding a sixth agent, an auditor, to catch what the reviewer misses, would be the same smell, with a new badge.

Why it happens

The metaphor is appealing. A team of agents is easy to explain, and looks good in a presentation. A loop with two tools and a rubric in code, much less so.

Frameworks encourage it. Many tutorials open with roles and goals written in HR language. You copy the example, swap the names, and in an afternoon you have five agents running. The demo talks. The cost shows up next month.

Splitting by human role, not by context. In a company, splitting by role makes sense, because people have memory, hours and a specialty. For a model, the reasons to split are different. The context window, tool permissions, review that mustn't see the maker's reasoning, and work that can run in parallel.

And no eval. Without measuring, nobody notices that three of the five agents change nothing. The reviewer improved the tone is a feeling. The rubric says whether it improved what matters.

What it costs

The cost. Each agent rereads the previous work, so cost grows with the number of roles, not with the value delivered. With illustrative numbers, the first call starts at three thousand tokens of input. Each later role receives the previous one's entire text. Five, seven, nine, eleven thousand. You pay for the same text several times, and what the extra buys is four rewrites. Prompt caching makes the repeated start cheaper. It doesn't erase the four extra calls.

And latency adds up too, because the roles run in a queue. Quality can get worse, like a game of telephone. A correct number in the researcher's summary becomes a different number in the writer's text. And when the report comes out wrong, there are five prompts to investigate, and nobody knows at which pass the fact changed.

The treatment

The treatment is a refactoring called collapse agents. Merge the agents that only pass text along into one agent, with the right tools. And keep separate only what has a technical reason. Those reasons do exist.

Context isolation. Whoever reads forty pages shouldn't write the report in that same context.

Restricted tools. Whoever summarizes a lead's website shouldn't hold the send email tool.

Independent review. Whoever scores doesn't see the writer's reasoning, and the eval shows a gain.

And parallelism, like twenty different repositories explored in twenty separate contexts.

Quick check. The eval shows that merging the page reader with the writer hurts the rubric. The context fills up with web pages, and the report loses its sources. The team still wants the five job titles from the slide deck. What's the most useful next step? Keep only the measured split, the reader and the writer, and collapse the rest. The eval proved one technical reason between two roles. Context isolation. Not five job titles.

And the classic trap, for someone starting out. Adding a reviewer, because a human team always has one. With no rubric and no eval, that reviewer learns to say approved. You paid for a stamp. And if the sequence of steps is always fixed, search, then write, then format, you may not need an agent at all. A chain, with code in the middle, does it, and each step has a contract.

Is it this smell?

Four quick cases. A pipeline with one agent that researches, and a second that writes from the sources, with no tools. And the eval shows that merging them hurts quality. Not this smell. The split has a measured technical reason.

Three agents debate in rounds. The optimist, the pessimist and the moderator. No criterion, no eval, and the final answer is usually the moderator's first one. That is this smell. A debate is only justified with a judge, a written criterion, and an eval showing improvement.

A main agent sends twenty subagents to explore twenty different repositories in parallel, and gets a one page summary from each. Not this smell. That's real, independent work, in separate contexts.

And after collapsing, you add an evaluator with a written rubric, that only sees the final report, and the score goes up. Not this smell either. That's independent review, with a measured gain.

Wrap-up

So here's the idea to take with you. A model doesn't need an org chart. It needs the right tools and a clean context. Drop each agent and measure. Keep only the splits that move the score, name their technical reason, and collapse the rest into one agent that does the work.

In the next lesson, we'll do that collapse step by step: collapse agents. See you there.