Multi-Agent Collaboration

When it's worth splitting into several, and where the added cost shows up.

"We use multiple agents working together" sounds advanced. But multi-agent isn't a higher mode — it's an account with real costs. Split right and you solve a problem; split wrong and you trade one hard problem for four.

Signs you'll recognize:

• A demo of a "multi-agent system" looks dazzling and you want to copy it.
• A single agent handles a task, and the context keeps overflowing.
• After splitting, you can't tell which step went wrong when debugging.

Why split at all

Splitting exists to fix three specific weaknesses of a single agent — and only those:

Context overflow. A single agent's entire memory lives in the context window. When a task gets long, parts get squeezed out. Split into several, each manages a small segment with its own window, and the problem is bypassed.

Role interference. Wanting both creative breadth and rigorous detail in one model, in one round, drags both down. Split, and one agent diverges while another converges, each with its own role prompt.

Parallelism or isolation. Several independent things need to run at once — or one step's permissions shouldn't be shared by the whole flow.

Turn it around: if a single agent isn't hitting these three, splitting is pure cost.

Three ways to split

Orchestrator. A "lead" agent understands the goal, splits the task, and delegates; several "worker" agents each do a piece and report back for consolidated decisions. Fits tasks that cleave cleanly and need unified control. The cost: the lead becomes a new bottleneck and single point of failure.

Pipeline. Strict order: A's output goes to B verbatim, B's to C. Each step has one role, can be evaluated in isolation, and is easy to debug. The cost: slowness — total latency is the sum of all steps, and one stuck step stalls the whole chain.

Peer. Several agents each do an independent task, then results are compared. Fits "generate several options, pick the best." Naturally parallel. The cost: who judges the final result, and by what standard, is often never settled.

The modes can be mixed. The key is deciding in advance who owns the final result. Without a clear owner, multi-agent is a meeting with no chair. OpenAI's agent guide covers these orchestration patterns in more detail and is a useful cross-check. openai-agents

Where the cost grows

The bill lands in four places:

Context shuttling. The lead copies task background to workers, and workers send results back. That content passes through billed input repeatedly. Split into N agents and total tokens often run several times higher than a single agent.

Who makes the call. Each agent sees only local information. Who has the authority for global judgment — and if it's wrong, whose fault? Untangle this or you get blame-shifting.

Harder attribution. When the final result is wrong, was it the task understanding, a worker's error, or information lost in hand-off? The search surface grows from one step to a whole chain.

Latency stacking. Serial chains sum the latency of every step, and users wait for the total. In interactive scenarios, multi-agent is often noticeably slower.

To split or not

A workable order:

First, optimize the single agent. Better prompt layering, tighter context, clearer tool docs. Most "a single agent isn't enough" is really "a single agent wasn't done well." Anthropic's guide gives the same advice: start with the simplest design and only add complexity when it's proven necessary. anthropic-agents

Second, confirm the split solves a specific problem. Window overflow? Role conflict? Real parallelism? Only if you can say "without splitting this can't work" do you earn the right to split.

Third, split the most painful step first. Don't turn it into five at once. Extract the one step that eats the most context or needs the most isolation, make it its own agent, keep everything else as is, verify the gain, then continue.

One standing discipline: every so often ask "can we split it back now?" Tools, models, and prompts all improve. A limit that today requires multi-agent may be handled by a single agent in six months. Multi-agent is a means, not the goal.