When is splitting actually worth it?

When the parts need genuinely different material in front of them, which is the benefit that survives scrutiny. A researcher and a writer working from different sources is a real split. A planner, a doer and a checker on one small task is usually one job wearing three hats and paying three bills.

What goes wrong that would not with one agent?

Handovers, mostly. Work passed between agents loses context in the way it does between teams, and nothing in the arrangement holds the overall goal. The result is a system where each part behaves reasonably and the whole drifts, which is harder to diagnose than a single agent going wrong.

Who holds the overall goal?

Usually something appointed to, and that appointment is the design decision people skip. Without it each agent optimises its own part and nothing notices the whole drifting. With it, the coordinator becomes the component whose failure takes everything with it, which is a real trade rather than a solved problem.

Usually several times more, because each agent runs its own loop and carries its own accumulated context. That is worth paying where the split is real. Where it is decorative it is an elaborate route to the answer a single agent would have produced, and the bill is the clearest signal of which one you have.

Glossary

Multi agent system

A multi agent system splits one job across several workers with different responsibilities, which helps on large tasks and introduces coordination problems that did not previously exist.

In plain terms

Instead of one worker doing everything, several take parts of it: one plans, one researches, one writes, one checks. It reads like an organisation chart, which is exactly why it appeals. It also inherits the problems organisations have, which is the half that gets less attention when the diagram is drawn.

Why it matters

Because the arrangement is easy to build and hard to judge. Frameworks make defining several agents almost as quick as defining one, so teams reach for it early, and the question of whether the job needed splitting rarely gets asked. The result is often a system that costs several times more and produces work of about the same quality.

How it works

The split is usually by responsibility, and the abstractions deliberately borrow from how people organise. Agents get a role, a goal and a set of tools, and are assembled into something that collaborates on a task. One framework in this guide describes its abstractions as reading like the org chart they imitate, and notes that this readability is the pedagogy: the arrangement is easy to reason about because it is familiar.

What it genuinely buys is focus. Each agent carries only what its part of the job requires, which keeps the material in front of it relevant and stops one enormous instruction trying to cover every case. That is a real benefit and it is about attention rather than about intelligence, which is why it helps most on jobs that are genuinely several jobs.

The coordination is the new problem, and it is not free. Something has to decide who does what, pass results along, and settle what happens when two agents disagree or one produces something the next cannot use. Handing work between agents is where information gets lost, in much the same way it does between teams.

Failure modes multiply rather than average out. The same framework's own documentation is blunt about it: unattended crews inherit every agent's failure modes at once, and the framework alone is not production governance. Each agent can be individually reasonable and the system as a whole still wander, because nobody is holding the overall goal.

Cost multiplies more straightforwardly than the benefit does. Every agent runs its own loop, resends its own accumulated context and bills for it, so a job split four ways is not a quarter of the work each. Where the split is genuine the extra cost buys something; where it is decorative it buys a more elaborate way to reach the same answer.

The diagram, and the bill

The org-chart resemblance is doing more work in this decision than anybody intends. A diagram with a planner, a researcher and a writer is immediately legible, matches how the work would be divided among people, and can be expressed in very little code, all of which makes it feel like the obvious design rather than one option among several. What the resemblance conceals is that the reason organisations split work that way is that people cannot hold everything at once, and that constraint does not transfer cleanly. Where it does transfer, because the parts genuinely need different material in front of them, the split earns its cost. Where it does not, the team has recreated an organisation's coordination overhead to solve a problem the organisation's structure existed to work around, and is paying a separate bill for each participant in the meeting.

Seen in the wild

A framework for role-based systems where agents get roles, goals and tools and are assembled into crews, with abstractions that read like the org chart they imitate.
CrewAI
A platform where multi-agent teams coordinate on bigger jobs, assembled from tools, triggers and instructions rather than code.
Relevance AI
An editor whose agent layer splits a task across parallel subagents, with plans reviewable before anything is built.
Cursor

Common misconceptions

People assume

More agents means more capability.

In fact

It means more coordination, which is a cost before it is a benefit. Splitting helps when the parts are genuinely different work needing different material in front of them. Splitting a single coherent task produces handovers, disagreements and several bills, and the answer at the end is frequently the one a single agent would have reached.

People assume

A framework that makes this easy means the arrangement is sound.

In fact

Ease of expression and soundness are unrelated, and the frameworks say so themselves. One notes plainly that it is not production governance, and that observability, reliability and controls are the platform work around it. Being able to draw the org chart in a few lines is not evidence that the organisation would function.

Telling them apart

Multi-agent system vs Subagent

Multi-agent system

Several agents with standing responsibilities, coordinating.

Subagent

One helper handed a narrow piece of a larger job.

A subagent is a delegation within one agent's work. A multi-agent system is an arrangement with its own coordination problem.

Questions

When is splitting actually worth it?: When the parts need genuinely different material in front of them, which is the benefit that survives scrutiny. A researcher and a writer working from different sources is a real split. A planner, a doer and a checker on one small task is usually one job wearing three hats and paying three bills.
What goes wrong that would not with one agent?: Handovers, mostly. Work passed between agents loses context in the way it does between teams, and nothing in the arrangement holds the overall goal. The result is a system where each part behaves reasonably and the whole drifts, which is harder to diagnose than a single agent going wrong.
Who holds the overall goal?: Usually something appointed to, and that appointment is the design decision people skip. Without it each agent optimises its own part and nothing notices the whole drifting. With it, the coordinator becomes the component whose failure takes everything with it, which is a real trade rather than a solved problem.
Does it cost more?: Usually several times more, because each agent runs its own loop and carries its own accumulated context. That is worth paying where the split is real. Where it is decorative it is an elaborate route to the answer a single agent would have produced, and the bill is the clearest signal of which one you have.

Key takeaways

The appeal is that it reads like an org chart; so do the problems.
What it genuinely buys is focus, keeping each agent's material relevant.
Coordination is a new cost: handovers, disagreements and no one holding the goal.
Failure modes multiply rather than average, which the frameworks say themselves.
Each agent runs its own loop, so cost multiplies more reliably than benefit.

Tools that use this

CrewAI
Roles, goals and tools assembled into crews, abstractions imitating an org chart.
Relevance AI
Multi-agent teams coordinating on bigger jobs, built without code.
Cursor
Parallel subagents splitting a task, with plans reviewable first.