Insight · September 6, 2026
Many agents,
one result.
One agent has a ceiling. Orchestration is the discipline of making several agents move as one, without losing the thread between them.
01 · The shift
The moment you run more than one agent, coordination becomes the job.
A single agent is already a system. It plans, calls tools, reads the result, and decides the next step. For a while, making that one agent better was the whole game. Then teams started running two, then five, then a shifting cast of them on the same task, and a new problem appeared that no single agent has.
Who decides what. What each agent is allowed to see. How the separate pieces come back together into one answer. What happens when one of them fails halfway through. That set of questions is orchestration. It is not a bigger model or a cleverer prompt. It is the wiring between agents, and it is where multi agent systems are won or lost.
02 · The problem
One mind is fast to reason with and slow to spread out.
A single agent holds one train of thought. That is its strength on a focused task and its ceiling on a wide one. Ask it to look into thirty companies at once and it works down the list, one after another, because it has one attention to spend. The task did not need one deep thinker. It needed thirty shallow ones working at the same time.
So you split the work across agents. And the instant you do, every part of the system that used to be free becomes something you have to design. The agents no longer share a memory. They can reach different conclusions from different context. Their outputs have to be reconciled by something. Split badly, a team of agents is slower and less coherent than the one agent you started with. That failure, not the wiring itself, is what orchestration exists to prevent.
The one line to keep
“A single agent has a train of thought. A team of agents needs a conductor.”
03 · What orchestration actually is
A handful of shapes, not an endless zoo.
In December 2024 Anthropic published Building Effective Agents, which sorts most coordination into a small set of building blocks. The names below are those blocks, read as orchestration. Almost every real system is one of them, or a few of them nested.
What are orchestrator workers and evaluator optimizer?
They are two of the workflows in Anthropic's Building Effective Agents. In orchestrator workers, a central model breaks a task down on the fly, delegates each piece to worker models, and synthesizes their results. In evaluator optimizer, one model generates a response while another evaluates it and gives feedback in a loop until it passes.
Source · Anthropic, Building Effective Agents, published December 19, 2024
01
Routing
A classifier reads the request and sends it to the agent built for it. One decision, no fan out. The lightest form of orchestration, and often the only one a system needs.
02
Sequential
One agent’s output becomes the next one’s input, along a fixed chain. Order is the whole design. It suits work with clear stages, where each step depends on the one before it.
03
Parallel
Several agents run at once on independent parts of the task, then a step merges what they return. Speed comes from the fact that the pieces never had to wait on each other.
04
Orchestrator worker
A lead agent reads the task, decides the subtasks on the fly, hands each to a worker, and synthesizes the results. The subtasks are not fixed in advance. This is the pattern behind most research and coding systems.
05
Evaluator optimizer
One agent produces, a second critiques the work against a bar, and the pair loops until it passes. The check lives outside the maker, so the system can hold a standard it could not hold alone.
04 · Not a headcount contest
More agents is not more intelligence.
It is tempting to read orchestration as scale. Two agents beat one, five beat two, so pile on more and the system gets smarter. It does not work that way. Each agent you add is another mind that can drift, another output to reconcile, another way for the run to fragment. Past a point, adding agents subtracts coherence.
Orchestration is a control problem, not a scaling one. The design question is never how many agents can I run. It is what is the smallest arrangement that gets this task done, who holds the decision, and how does the work come back together. A routing step with two specialists can be better engineering than a swarm of ten. The art is choosing the smallest shape that fits, then instrumenting it so you can see what actually happened.
05 · One task that shows the difference
A question that is wide, not deep.
In June 2025 Anthropic described the orchestration behind its Research feature. A lead agent reads the request, plans an approach, and spins up three to five subagents that search in parallel, then a separate pass checks every citation before the answer is assembled. They gave one task that exposes the whole point. List the board members of every company in the Information Technology arm of the S&P 500.
A single agent grinds through that request one company at a time, because it has one line of attention. The orchestrated system splits it, sends a worker after each slice at the same time, and gathers the results. On Anthropic's internal research eval, the multi agent setup, a Claude Opus 4 lead with Claude Sonnet 4 workers, outperformed a single agent Claude Opus 4 by 90.2 percent. The gain was not a smarter answer. It was coverage a single train of thought could not reach in the time.
lead read the task · this one is wide, not deep plan one slice per company, no dependencies spawn 3 to 5 workers search their slices at once gather collect findings as each worker returns cite a separate pass verifies every source answer synthesize only once the branches are in
An orchestrator worker plan · breadth the lead could not walk alone
06 · When it pays off
The win is real, and so is the bill.
The same Anthropic writeup that reports the 90.2 percent gain also reports its price. That research system uses about fifteen times the tokens of an ordinary chat, and the team found that token spend alone explained roughly eighty percent of the difference in how the systems scored. Orchestration bought the result, and it bought it with compute. That trade is worth making on work where breadth or reliability is the point, and wasteful on work where it is not.
So the honest test is the shape of the task. Reach for orchestration when the work fans out into independent parts, when it spans sources a single pass cannot hold, or when a wrong answer is expensive enough to justify a second agent checking the first. On anything narrow, the extra agents add latency and cost for a result one agent already reaches. This is also why observability matters more the moment you orchestrate. You cannot manage a system of agents you cannot see.
07 · When not to reach for it
Dispersed decisions are how these systems break.
The sharpest warning comes from the other side of the debate. In 2025 Cognition published Don't Build Multi-Agents, arguing that when you fan work out to agents running on their own, the decisions get dispersed and the context fails to travel between them. Each agent acts on a partial view, they drift out of sync, and the final reconciliation cannot repair what they never agreed on. Their guidance is to share context as widely as possible and keep the actual writes on a single thread.
Both camps are describing the same truth from opposite ends. Orchestration pays when the work is genuinely parallel and the pieces do not need to agree. It fails when a task is really one connected line of reasoning that got chopped across agents who could not see each other. The reflex to avoid is spinning up a swarm because it feels advanced. Most tasks want one strong agent with the right context, not a committee. Orchestrate when the shape of the work leaves you no simpler choice.
Closing
Many agents, one result.
Take the task where your one agent keeps running out of attention, the wide one with parts that never had to wait on each other. Give it the smallest arrangement that fits, one lead and a few workers, and watch what comes back. The first time coverage arrives that no single thread could reach, you will start designing the wiring, not the agent.
Written by · September 6, 2026
Donny Smith · ECD, Founder, Bttr.
Over the past 15 years, he has led creative teams and contributed to products used by millions of people worldwide, working with companies including Apple, Opendoor, JP Morgan, GE Aerospace, Pepsi, and Alterra Mountain Company.
LinkedInAnthropic, Building Effective Agents, December 2024 · Anthropic, How We Built Our Multi-Agent Research System, June 2025 · Cognition, Don't Build Multi-Agents, 2025
Share this perspective
More insights
Adjacent perspectives.
Bttr. Field Brief
The brief Bttr. writes for senior buyers.
Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.