The multi-agent tax, itemized
Lesson 3 of 5 in Multi-Agent Systems: When One Agent Isn’t Enough.
Every agent you add bills you four ways, and only one of them shows up on an invoice.
Token multiplication. Each agent re-sends its own system prompt, tool schemas, and brief on every call; briefs must be written, reports must be written and then re-read by whoever integrates them. Coordination text is pure overhead — tokens spent on talking about the work instead of doing it.
Latency stacking. In serial shapes — handoff chains, debate rounds, any supervisor step that waits on a worker — each boundary adds full model round-trips to the critical path. Parallel fan-out is the only shape that refunds latency, and only when the workers are genuinely independent.
Here is the honest arithmetic for one task — “compare five competitors’ pricing pages, produce a two-page briefing” — done both ways. The numbers are illustrative bookkeeping, not a benchmark; the structure of where they come from is what transfers.
| Line item | Single agent | Supervisor + 5 workers |
|---|---|---|
Model calls | 14 — ~2–3 tool steps per competitor, then synthesis | 36 — supervisor 6, workers 5 × 6 |
Fixed overhead re-sent every call | ~2k tokens × 14 calls ≈ 28k | ~2k tokens × 36 calls ≈ 72k — six system prompts and toolsets instead of one |
Input tokens | ~225k — one context growing 3k → 30k, re-sent each call | ~300k — five fresh worker contexts (isolation keeps each small) + briefs + reports re-read by the supervisor |
Output tokens | ~8k — the briefing | ~17k — a plan, five worker reports, and the briefing |
Token bill | 1× | ~1.4× here — the friendly case. Vague briefs, duplicated exploration, or adding a debate stage push real systems to several-× more |
Wall clock | ~14 serial round-trips | ~12 round-trip equivalents: plan (2) + slowest worker (6) + integrate (4). Parallelism roughly breaks even here and wins as per-worker work grows |
Contexts to search when the answer is wrong | 1 | 7 — plus ten briefs and reports flowing between them |
Compounding non-determinism. Each inter-agent boundary adds sampled decisions of a new kind: what to put in the brief, what to cut from the report. Every hop is a lossy, non-deterministic compression. If each hop preserves the task’s intent 95% of the time, a five-hop chain delivers intact intent about 77% of the time (0.95⁵) — and unlike a failed tool call, a subtly warped brief throws no error. It just produces confident work on the wrong task.
The debugging problem. The final memo says a competitor cut prices 40% — the real number is 4%. Whose context held the mistake? Worker 3’s raw fetch? The compression into its report? The supervisor’s synthesis over five reports? With one agent you scroll one transcript. With seven contexts, you need traces that stitch every agent’s spans into one tree, or the error’s birthplace is simply gone — most frameworks discard worker contexts the moment they return.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.