When one agent isn’t enough: give it an org chart
Four measurable symptoms of an overloaded agent, and why the answer is an organisation rather than a bigger brain.
The control layer is built and it works. Now the business asks for more: this ticket also needs somebody to work out who else was affected, and somebody to draft what we say, and both of those need doing before a proposal is worth reviewing.
The reflex is to add tools and paragraphs to the agent you already have. There is a point past which that stops working, and it has recognisable symptoms.
| Symptom | What is actually happening | The role that fixes it |
|---|---|---|
| an instruction sheet nobody can audit | four jobs are interleaved in one prompt, so no sentence can be changed safely | one agent per decision, each with its own short brief |
| it confuses tools that look alike | the tool list has grown past what a single call can discriminate between | give each specialist only its own tools |
| it runs out of iterations before finishing | one oversized intermediate result is being paged through the context repeatedly | a specialist returns a conclusion, not a transcript |
| answers drift between runs | the agent is deciding what to prioritise as well as what to answer | a manager whose only job is to ask the right specialist the right question |
an instruction sheet nobody can audit
- What is actually happening
- four jobs are interleaved in one prompt, so no sentence can be changed safely
- The role that fixes it
- one agent per decision, each with its own short brief
it confuses tools that look alike
- What is actually happening
- the tool list has grown past what a single call can discriminate between
- The role that fixes it
- give each specialist only its own tools
it runs out of iterations before finishing
- What is actually happening
- one oversized intermediate result is being paged through the context repeatedly
- The role that fixes it
- a specialist returns a conclusion, not a transcript
answers drift between runs
- What is actually happening
- the agent is deciding what to prioritise as well as what to answer
- The role that fixes it
- a manager whose only job is to ask the right specialist the right question
Seen in this build. The manager agent spent six of its roughly ten available iterations paging through one specialist’s oversized answer. That is what “hit the iteration limit before it finished” looks like from the inside, and it is a design problem — the specialist was returning everything it had read instead of what it had concluded.
An organisation, not a bigger brain
Notice what has not changed: the reviewer, the gate, the human approval and the audit row are exactly the same nodes as in module 4. The team replaces the proposer. It does not replace the control layer, and adding agents to a system with no control layer multiplies the problem rather than solving it.
Do not reach for a team first. Every specialist is another model call, another place a fact can be invented, and another prompt to maintain. Split an agent when you can name the symptom above that you are suffering from — not because an org chart looks more serious than a single node.
Related lessons
Base rates — what a piece of evidence is actually worth
A face-recognition system that is 99.9% accurate and almost entirely wrong, and a number that sent an innocent woman to prison. Both are the same arithmetic, and it is the arithmetic that decides what any piece of evidence is worth.
ReadConfirmation and survivorship — what you never looked for
Two questions about evidence you did not go looking for. One is a rule you have to discover, and one is a pattern in five famous people — and in both, the thing that would have told you the truth is the thing nobody checks.
ReadLoss aversion, sunk cost and regression — what it costs you
Four questions you answer about yourself rather than about a scenario, and your own answers are the finding. Then the pattern that makes praise look useless and criticism look like it works, whatever you actually do.
Read