Module 075 min

Many agents: power or paperwork

Multiple agents sound like a team. Usually they're a meeting. Here's when splitting the work genuinely helps — and what it costs.

After this module
  • You can tell the three legitimate reasons to add a second agent from the two reasons people usually do.
  • You know the failure modes that only appear once agents talk to each other.

A multi-agent system is just several loops, each with its own context, coordinating through messages or a shared workspace. Nothing new appears mechanically. What changes is the economics: every extra agent adds a context, a model bill, a latency chain, and a new way to be wrong.

7.1Three reasons that hold up

1 · Context isolation

A sub-agent that reads forty files and returns a one-page summary keeps the main agent's context clean. This is the strongest reason, and it applies to almost every coding or research agent.

2 · Genuine parallelism

Five independent sources, twenty independent files, three independent candidates. Wall-clock time drops roughly with the count — if the work really is independent.

3 · Different tools or permissions

A worker with database access and no internet, and a worker with internet and no database. Separation of capability is a safety boundary, not just a design taste.

The two reasons people usually give — "specialised personas will debate and reach a better answer" and "more agents means more intelligence" — are the ones to be suspicious of. Debate between models that share the same training data and the same context produces agreement, not truth. And capability doesn't multiply; it only gets redistributed.

THREE TOPOLOGIES SUPERVISOR Lead Worker 1 Worker 2 clear ownership HIERARCHY Manager Team lead Team lead w1 w2 w3 scales, adds relays PEER SWARM A B C D emergent · hard to audit
Every layer you add is a place a message can be lost. Supervisors are debuggable. Hierarchies scale but relay information through more mouths. Swarms are interesting research and a poor default for production.

7.2Delegation vs handoff

Two different moves, often conflated:

Delegation
The parent keeps the goal and spawns a worker for a bounded piece of it. The worker returns a result and ends. The parent remains the owner of the outcome. Use for: search this codebase, summarize these papers, run these tests.
Handoff
Ownership moves. The conversation becomes the specialist's, and the parent is no longer driving. Use for: "this is a billing issue, the billing agent takes it from here".

The practical rule: delegate work, hand off conversations. Most bugs come from mixing the two — a worker that keeps talking to the user, or a handoff that leaves the original agent still trying to answer.

7.3Failure modes that only exist here

The telephone game
Each hop rewrites the summary. By agent four, the constraint "must be under $2,000" has become "reasonable budget". Fix: pass structured data (JSON with fields), not prose, between agents; make the original goal travel with every sub-task.
Cost multiplication
Five agents × eight steps × a 40k context is forty model calls and 1.6 million input tokens for one task. Multi-agent systems routinely cost 5–15× a single-agent equivalent. Measure cost per completed task, not per call.
Deadlock and livelock
Two agents waiting for each other's output, or politely passing a task back and forth forever. Fix: explicit ownership, timeouts on every wait, and a hard rule that a task may be reassigned at most twice.
Silent divergence
Workers each produce a plausible answer and the merge step picks one without noticing they contradict. Fix: make the merge step compare answers and surface conflicts instead of averaging them into a confident paragraph.
Try it · which shape for which task?four tasks

The default that works

One agent, better tools, explicit state. Add a second agent only when you can name which of the three reasons applies and measure the improvement. If you can't state the reason in one sentence, the second agent is architecture theatre.

Check yourselfQuestion 1 of 1

Which is the strongest justification for adding a sub-agent?

60-second recap

  • Multi-agent is an economics decision, not an intelligence one: contexts, bills, latency and failure modes all multiply.
  • Three legitimate reasons: context isolation, real parallelism, different tools or permissions.
  • Delegate work, hand off conversations. Keep the owner of the outcome unambiguous.
  • Pass structured data between agents — prose summaries degrade with every hop.
  • Start with one agent. Earn the second one.