Wiki
Core10 min read

Why many agents

Splitting work across agents buys fresh contexts and specialisation — and pays for it in coordination. The topology decides which you get more of.

One agent has one context window, and the previous chapter showed that window is the bottleneck. Multi-agent systems are, at heart, a way to buy more context: give each subtask its own agent with its own clean window, and let them coordinate. The benefit is real; so is the cost, and the topology you choose sets the exchange rate.

Start here

An extra agent is an extra context window — which is exactly what you want when one window cannot hold the whole task. It is also an extra thing to coordinate. Multi-agent is not "more intelligence"; it is "more context, plus a coordination bill."

Pick a topology — more agents, more coordination

agent
agent 1
agent 2
agent 3

agents

4

coordination links

3

A lead splits the task and dispatches subtasks to workers with fresh contexts. Scales context; adds a merge step.

Every added agent is a fresh context (good) and a new coordination link (costly). Coordination grows faster than agents — pairwise chatter scales quadratically — which is why most successful "multi-agent" systems are really one orchestrator over a few workers, not a crowd. Figures are illustrative.

Switch topologies and watch the agent count and coordination links move together.

Why split

  • Context isolation — a research sub-agent reads 50 documents into its window and returns three sentences, so the lead's window stays clean.
  • Specialisation — different prompts, tools, or even models per role: a cheap router, a careful reasoner, a strict verifier.
  • Parallelism — independent subtasks run at once, cutting wall-clock time.

The coordination tax

Everything you split, you must recombine. Results have to be merged, disagreements resolved, and shared state kept consistent. Coordination grows faster than the agent count — pairwise communication is quadratic — so the marginal agent buys less context and more overhead than the one before it.

Careful

"Add another agent" is the multi-agent version of "add another if-statement". Most tasks that look like they need a team are better served by one agent with better tools and a tighter context. Split when a subtask genuinely needs its own window or its own model — not because the diagram looks impressive.

Check yourself

Eduspheria wiki · Agentic AI, Multi-agent & orchestration

0 / 4 answered

  1. 1What is the fundamental reason to split work across agents?
    Multiple choice
  2. 2If coordination links grow with the square of the agent count, how many pairwise links are there among 5 agents?
    Numeric answer
  3. 3The marginal agent buys more context and less overhead than the agent before it.
    True / false
  4. 4What resource do multiple agents buy, in the lesson's framing?
    Short answer

Next: the one multi-agent shape that earns its keep most often — a lead over workers.