Multi-Agent Context Sharing

Separate windows are the point

A multi-agent system is usually described as a way to divide labour. The context argument is stronger: each subagent has its own window, so work that would have competed for one budget now runs in several.

Anthropic describes the effect as compression. Subagents operate "in parallel with their own context windows, exploring different aspects of the question simultaneously before condensing the most important tokens for the lead research agent."

That last clause is the mechanism. The lead never sees the forty pages each subagent read. It sees what each one concluded.

The orchestrator-worker shape

One lead agent analyses the request, forms a strategy, and spawns subagents to work on parts of it at the same time. They gather, they condense, they report back, and the lead assembles the answer.

The context property that makes this work: a subagent's window fills with material the lead never has to carry. A research task that would blow a single window finishes in five smaller ones.

Context isolation covers giving a sub-task its own window as a general technique. This page is what happens when there are several, and they have to agree.

What a subagent has to be told

The delegation is the whole game, because a subagent cannot ask a clarifying question of the thing that spawned it.

Anthropic's list is four items: an objective, an output format, guidance on the tools and sources to use, and clear task boundaries.

Boundaries are the one teams leave out, and the failure is specific. Without them, subagents duplicate each other's searches, because two agents given overlapping objectives will both reach for the obvious source. You pay twice for one result.

Match the specificity to the task. A simple query wants one agent making three to ten tool calls. Complex research can justify more than ten subagents, but only with responsibilities divided so that no two overlap.

The cost is the headline

The multipliers are large and worth quoting exactly: agents use about 4× the tokens of a chat interaction, and multi-agent systems about 15×.

Fifteen times. That is the number to hold up against whatever the architecture is supposed to buy, and it is the reason this pattern belongs to tasks where the answer is worth real money. Anthropic's own justification is a research eval where a multi-agent configuration beat single-agent Opus 4 by 90.2%, which is the kind of margin that pays for a 15× bill.

For a task where a single agent gets most of the way, it does not.

How delegation fails

The documented early failures are instructive because none of them is the architecture being wrong.

Spawning fifty subagents for a simple query. The lead has no cost intuition unless you give it one. Bound the fan-out explicitly.

Searching for sources that do not exist. A subagent with an objective and no stopping condition will keep going. Say when to give up.

Duplicated work from vague instructions. Covered above, and the cheapest one to fix.

The pattern across all three: a subagent does exactly what its task description says, and a vague description produces expensive nonsense rather than an error. Write the description as though it is going to a contractor who cannot phone you, because that is the situation.

What does not cross between windows

Two things to design around.

Subagents cannot see each other. Anything two of them need is either in both task descriptions or in shared storage they can both read. External memory is the usual answer when the shared thing is large or accumulates during the run.

The lead sees only what comes back. A subagent that found something important and did not include it in its report might as well not have found it, which is why the output format is one of the four required items rather than a nicety.

Further reading

• Anthropic, How we built our multi-agent research system: the pattern, the four-part task description, the token multipliers and the early failures.
• Context isolation: the single-window version of the same argument.
• External memory: where agents that cannot see each other put shared state.
• Multi-agent systems: the shapes this is one of.