The Swarm Is a Different Animal

One agent is a hire. Five agents are a team, and teams have team problems. What the 90 percent swarm gain costs, and the shape of work where it appears.

Alaa Almallah June 5, 2026 9 min read

One agent is a hire. You brief it, it works, you read the result. Most teams stop there and most of the time that is enough.

Then someone adds a second agent, and a third, and a strange thing happens: the work does not get three times better. Some of it gets dramatically better, some gets worse, and all of it gets more expensive. The team has discovered the swarm, and the swarm is not a bigger version of the hire. It is a team, with everything that makes teams powerful and everything that makes them difficult.

The number that started the conversation

Anthropic described their research system in 2025: a lead agent that plans a query, then spawns three to five subagents that search in parallel, each with its own context window, compressing findings back for the lead. On their internal research evaluations, that system beat a single-agent setup by 90.2 percent.

The same document carries the bill. The multi-agent system used roughly fifteen times the tokens of a single chat to do it. Both numbers are the number. A swarm is a trade: you pay coordination and compute to buy coverage and parallel judgment.

When the swarm wins

The 90 percent was not uniform. The gains concentrated where the work has a particular shape:

Breadth-first tasks. Questions answered by covering a lot of ground independently: research across many sources, scanning many documents, exploring many alternatives. Parallel coverage beats one careful reader when the space is wide and the pieces do not depend on each other.
Verifiable pieces. Each subagent returns something that can be checked against a standard, so the lead merges evidence rather than opinions.
Clean interfaces. Each agent owns a slice with a known boundary, and the slices hand back compressed results, not open-ended thoughts.

Product work often has this shape and nobody planned it that way. Competitive scans, test-matrix generation, option exploration, literature review: these are swarm-shaped, and teams leaving them single-threaded are leaving the 90 percent on the table.

When the swarm fails

The same system struggled exactly where the shape breaks:

Shared context. When every piece needs the whole picture to make sense, subagents working blind produce confidently incompatible results, and the lead spends its budget arbitrating instead of synthesizing.
Sequential judgment. When step two's meaning depends on step one's outcome, parallelism is fake. You are just paying fifteen times the tokens to run a queue.
Mutable shared state. Two agents editing the same thing is not a team. It is a collision with extra steps.

The deep reason is the one this series already argued: agents should not share a brain. Separation is what makes parallel judgment honest. But separation has a price too, and the price is that somebody, the lead agent or a human, must own the interfaces between the pieces. A swarm without owned interfaces is five strangers in a room, each certain, none responsible.

The real cost is coordination

The token bill gets quoted because it is easy to count. The harder cost is coordination, and it arrives in three forms:

Decomposition cost. Someone, human or lead agent, must split the task honestly. A bad split produces five beautiful halves that do not compose.
Merge cost. Subagents return compressed findings, and compression means judgment: what was dropped, what was weighted, what contradicts. The merge is where errors hide, because each piece looked fine locally.
Failure-mode variety. One agent fails loudly and obviously. A swarm fails quietly and partially: one subagent's weak search silently narrows the whole answer, and nothing beeps.

This is why the swarm needs exactly what every team needs, a contract layer and an escalation path, just enforced between machines instead of between colleagues. The teams doing well with swarms did not get better agents. They got better at writing the boundaries.

A quiet rule of thumb

Before spawning a second agent, answer three questions in writing:

Can this task be split so the pieces are independent, or am I faking parallelism on a sequential job?
What standard will each piece be checked against, and who wrote it?
Who merges, and do they know the merge is where the errors will hide?

Three real answers mean the swarm will earn its fifteen times. A shrug at any of them means the single agent, run twice with better briefs, is the cheaper animal this month.

FAQ

Is the 90 percent number credible if the vendor published it?

Treat it as an honest vendor measurement with a scope: internal research evaluations, breadth-first queries, their own models, 2025. The direction is corroborated by the shape of the task, and the shape is what matters. Test the pattern on your own breadth-first work and keep your own number, dated.

Does the 15x token cost not make swarms pointless?

It makes them a budget decision, which is healthier than a fashion. With unit prices falling roughly ten times a year for constant capability, a15x bill today costs what a 1.5x bill cost recently. The question is whether the task returns fifteen times the value. Broad research often does. Routine drafting never will.

Should every team learn orchestration now?

Teams should learn to recognize swarm-shaped work, which is different from orchestrating everything. The skill is seeing which tasks are breadth-first and verifiable, then writing clean interfaces for the split. If none of your current work has that shape, the highest-leverage move is a better brief to one agent, not a smaller fleet.

How does this connect to human teams?

Directly. A swarm is a small organization with the same physics: clear ownership of slices, standards for what comes back, an owner of the merge, and an escalation path when a piece is stuck. Managers will recognize the job. The swarm did not invent team problems. It borrowed them.

If you are deciding which work deserves a swarm and which deserves one good agent, book a discovery call.

Part 7 of 14 · The Fleet Series

From one person steering ten agents to hierarchies, routers, and the executive table: what multiplies, what it costs, and where it breaks.

Previous in the series One Minute, One Hundred Calls Next in the series The Unit Price Falls While the Bill Grows

Advanced Venture Building