Anthropic published a guide on when to use multiple AI agents and when not to. Read it as someone who builds councils and you reach an uncomfortable conclusion: by its rules, a council is the thing you should never build.
The guide names three situations where multi-agent wins — context isolation, parallel execution, specialization — and one where it loses: tasks where every agent needs the same context and the steps depend on each other. "Domains that require all agents to share the same context... are not a good fit for multi-agent systems today."
That last line is a precise description of a council. Same question. Same context. Every model in the room looking at the identical problem. By Anthropic's criteria, that's the anti-pattern.
The lab that proved multi-agent works is telling you where it stops
This isn't a stray sentence. Anthropic is the company that built the lead-agent-plus-subagents pattern and reported it beating a single Claude Opus by 90.2% on internal research evaluations. When the lab that demonstrated multi-agent at scale also draws the boundary around it, builders should listen.
And the boundary it draws — shared context, tightly coupled, many dependencies between agents — is exactly where a decision council lives. So either the council is a mistake, or Anthropic is answering a different question than the one a council answers.
It's the second. Anthropic's guide is about dividing labor. A council is about distrusting a verdict. One splits a task so it fits. The other duplicates a task so it gets checked.
All three winners are throughput moves
Look at what each of Anthropic's three scenarios actually optimizes. Context isolation keeps a subagent's research from polluting the lead agent's window. Parallel execution runs ten searches at once instead of in sequence. Specialization hands the SQL to the SQL agent and the writing to the writer. Every one of them is about getting more done per unit of time. None of them is about being wrong.
A council does the opposite of dividing the work. It hands every model the whole job, on purpose, because the point isn't speed — it's catching the confident wrong answer that a single model would have given you with a straight face. The two designs aren't competing. They're shaped for different failures.
Decomposition optimizes for coverage. A council optimizes for disagreement.
When Anthropic splits a research task across subagents, success looks like coverage — did the swarm find everything? The subagents are not supposed to overlap. Overlap is waste. That's why shared context is a bad fit: if two agents are doing the same thing, you're paying twice for one result.
A council inverts the metric. Overlap is the product. You run Claude, GPT-5, and Gemini on the same question precisely so you can see where they diverge. Ask one model whether a contract clause is enforceable and you get a clean, confident yes. Ask three and watch one of them flag a jurisdiction problem the other two sailed past. The disagreement is the signal. In a decomposition system, that same disagreement would be a bug to engineer away.
"Dependencies between agents" is a cost in one design and the whole mechanism in the other.
Anthropic warns off tasks with many dependencies between agents, and it's right to — in task decomposition, every dependency is a coordination tax. Agent B waits on Agent A. The lead has to reconcile. Latency compounds and the parallelism you bought evaporates.
A council's dependencies aren't coordination. They're confrontation. Red Team vs Blue Team is nothing but dependency — one side exists to attack the other's answer. Collaborative Editing is models building on and cutting into each other's drafts. The dependency is the work, not the overhead. Measured against a throughput goal it looks like pure cost. Measured against "don't ship a confident error," it's the only part that pays.
The shared context Anthropic avoids is the shared context a council requires.
Here's the cleanest way to see the gap. Anthropic's subagents are deliberately blind to each other — in its own words, a subagent "doesn't know the other subagents exist." That blindness is what makes parallelism safe and keeps the lead agent's window from drowning in cross-talk.
A council can't be blind. Every member has to see the same question, and the synthesis layer — the Chairman — has to see every member's answer to decide what to do with the disagreement. Shared context isn't a fit problem to route around. It's the substrate the whole thing runs on. Strip it out and you don't have a leaner council. You have no council.
Two right answers in different languages
Anthropic's guide is right, and the council is right, and they don't contradict — they're sentences in different languages. If your problem is "this task is too big for one context window," follow Anthropic's rules to the letter: isolate, parallelize, specialize, and keep the agents out of each other's way. That advice is hard-won and correct.
If your problem is "this decision is too expensive to get wrong," throw those rules out and do the forbidden thing — same question, same context, every model in the room, and a synthesis layer whose only job is to decide what the disagreement means.
The tell is in what you're optimizing. Decomposition asks: did we cover it? A council asks: should we trust it? Anthropic answered the first question well. The second one is still sitting on your desk.
Try it free — no signup. shingik.ai