Adding agents made our pipeline slower and more expensive, and the research now says why. In July 2026 Nature Machine Intelligence published a controlled study titled "Capable language models can outgrow the benefits of collaboration." The authors held task prompts, tools and compute budgets constant and varied only the coordination structure. Multi-agent systems incurred 58% (independent agents), 263% (decentralized), 285% (centralized) and 515% (hybrid) more reasoning turns than a single agent, which is 1.6x to 6.2x the work for the same job. At equal token budgets, the single agent matched or beat the crews on reasoning tasks. The strongest predictor of whether coordination helped at all was the single-agent baseline: when the base model was already good, adding agents mostly added cost.
Short answer: default to one agent. Allow a second only when a task passes one of three tests. Replace the rest of the crew with shared state (locks), a gate (QA), and a budget. The 2026 story for solo builders is orchestration discipline, not headcount.
What the study measured, and what it did not
The paper is not an argument that multi-agent systems never work. It found two conditions where coordination helps: work that is genuinely parallelisable and read-heavy, and narrow domains where an extra checker improves reliability. Outside those, the overhead is pure. The important design point is the phrase "under matched per-system compute ceilings." Most crew demos never match compute. They give the crew ten times the tokens and then celebrate a slightly better answer.
For a one-person studio the framing is simple. You are paying for every reasoning turn. A pattern that multiplies turns by six to produce the same artefact is a tax. If you have read our piece on the agentic cost crisis, you already know the enterprise version of this bill. The solo version is smaller in dollars and larger in proportion.
The one-agent default law
Write it down where the team, or the future you, can see it: every task starts as one agent with a written brief, a tool list and a token cap. Nobody spins up a Director, a Researcher, a Writer and a Critic because the framework template had four boxes. The burden of proof is on the second agent.
This sounds obvious and is violated constantly, because frameworks make crews the path of least resistance. A single agent with a good brief, a project bible it can read, and a verification step it must pass will out-produce a chatty crew on most creative and coding tasks. The Nature MI result is the quantitative version of what every operator who has watched five agents agree with each other already knew.
The three tests that allow a second agent
| Test | Question | If yes |
|---|
| Different capability | Does the subtask need a different model, a different tool set, or a different context window than the main agent has? | Split. A vision model reviewing frames is not the same worker as the script writer. |
| Audit wall | Must the checker be unable to see the maker's reasoning, so that it cannot be talked into approving? | Split. A QA gate that reads only the artefact and the spec is worth its tokens. |
| Real parallelism | Are there N independent, read-heavy units of work that finish faster side by side and do not need to agree with each other? | Fan out, then merge with one agent. This is the parallelisable case the paper identified. |
Fail all three and you do not get a second agent. You get a better brief. "The writer needs a critic" fails the audit-wall test if the critic shares the writer's transcript, because it will simply confirm the writer. "Research then write" fails the capability test if the same model with the same tools does both. Most crews we have audited fail all three on every hop.
Locks instead of agents
The work people assign to extra agents is usually memory. A Continuity agent exists to remember what the last shot looked like. A Style agent exists to remember the voice. That is state, and state belongs in a file, not in a personality.
We run two documents that replace three agents:
Consistency Lock: the immutable facts of the project. Names, palette, voice rules, forbidden words, output formats. Read at the start of every run. Never regenerated.
Identity Lock: for anything with a recurring character or brand: the traits that may not change, the traits that may, and the drift checks. The video version is covered in Identity Lock for AI video.
A lock costs a few hundred tokens to read. An agent whose job is to remember the lock costs a full reasoning turn every time it speaks, plus the turns everyone else spends replying to it.
The gate: a QA guardian that does not chat
The one second agent we always allow is the gate, because it passes the audit-wall test. It reads the artefact and the spec. It does not read the maker's transcript. It returns a structured verdict, not a conversation.
{
"artifact": "scene_04.mp4",
"checks": [
{ "id": "identity_face", "pass": true },
{ "id": "identity_wardrobe","pass": false, "note": "jacket colour drifted to navy" },
{ "id": "continuity_light", "pass": true },
{ "id": "spec_duration", "pass": true },
{ "id": "spec_aspect", "pass": true }
],
"verdict": "REJECT",
"reject_class": "wardrobe_drift",
"retry_allowed": true
}
Sixteen checks is our ceiling. Past that the gate becomes a second author. The output is machine-readable so that the orchestrator, not another agent, decides what happens next: retry with the lock re-injected, escalate to a human, or ship.
Drift score: when to kill a run
Crews fail slowly. Each agent nudges the artefact a little, nobody owns the total, and by turn twelve you are polishing something that no longer matches the brief. Give every run a drift score: the number of gate checks that failed across the last three attempts. Two consecutive failures on the same check means the fix is in the brief or the lock, not in another retry. Kill the run, fix the document, start clean. This single rule cut our wasted generations more than any model upgrade.
Token budget per finished artefact
Budget the artefact, not the session. A finished 8-second clip, a shipped blog post, a merged pull request: each gets a token ceiling derived from what it earns or saves. If a pattern cannot deliver inside the ceiling, the pattern is wrong. Use our free AI token estimator to size briefs and locks before you run them, and the prompt cost calculator to turn a turn count into money per model. When a crew's cost per accepted artefact is six times a single agent's, the Nature MI numbers stop being academic.
Migrating a crew to an orchestrator in one afternoon
List every agent in the current crew and the one sentence it exists for.
Any agent whose sentence contains "remember", "keep consistent" or "make sure the style" becomes a lock file. Delete the agent.
Any agent whose sentence is "review" or "critique" and which can see the maker's transcript becomes the gate, with the transcript removed and a JSON verdict added.
Any agent whose sentence is a step in a sequence ("research, then outline, then draft") collapses into the main agent with the steps written into the brief.
Whatever is left must pass one of the three tests. Usually one does, sometimes none.
Add a token ceiling and a drift rule. Run the same task on the old crew and the new orchestrator with the same budget. Keep the winner.
We have not yet seen a solo pipeline lose that comparison.
Where skills fit
The thing that replaces most of a crew's specialists is not another agent. It is a written skill: a reusable procedure with inputs, forbidden actions and a verification step that the single agent loads when it needs it. That is the subject of Agent skills beat agent crews. If you want the locks, gate schema and skill templates as files rather than prose, the AI Agent Ops Bundle ships the spec, observability and cost-control set we run, and the Agent Prompt Vault has the brief templates. For measuring the gate in production, read the agent observability guide.
Quick answers
Did the Nature Machine Intelligence study say multi-agent systems are useless?
No. It found they help for parallelisable, read-heavy work and for narrow-domain reliability. Elsewhere they added 1.6x to 6.2x reasoning turns with no accuracy gain at matched compute.
Is a critic agent ever worth it?
Yes, when it cannot see the maker's reasoning and returns a structured verdict. A critic that shares the transcript tends to agree with the maker and costs a full turn to do so.
How do I know my single agent is strong enough?
Run the task once with a proper brief and lock and measure the gate pass rate. The paper found the single-agent baseline is the best predictor of whether coordination will help.
What replaces the memory agents?
Lock files read at the start of each run: a Consistency Lock for project facts and an Identity Lock for recurring characters or brand traits.
What is a sane token budget?
Derive it per finished artefact from what the artefact earns or saves. Size briefs with a token estimator before running and compare cost per accepted output, not cost per generation.
Fewer agents, better documents, one gate. Every product mentioned is available at wowhow.cloud — pay once, ship forever.
Originally published at wowhow.cloud
Top comments (0)