“Most of the time, honestly. A multi-agent system is a context-management strategy: workers burn their own windows on the reading and hand the lead a short report. You pay for that in tokens — Anthropic measured roughly four times a chat for one agent and fifteen times for a team — so it only wins when the work is broad, parallel and worth that bill.
It’s worse when the steps are coupled. Every action carries implicit decisions, so two workers editing the same module make conflicting ones, and the merge is worse than either. It’s worse when the task fits in one window anyway, and worse when the task is cheap, because a router in front would have sent it down a one-call path.
It’s also worse to operate. Every hop loses context, results come back partial, retries multiply per layer, and a run that says ‘success’ at the root can have two failed children under it.
So I start with one agent and only split when an eval shows context pollution or tool confusion in it — and then a star, two levels, parallel reads, serialised writes.”
The deep dive — diagrams, tradeoff tables, and the follow-up trap
Two questions decide it: is the work separable, and is it worth 15×?
Multi-agent wins in one quadrant: high-value work that splits into independent reads
Where a team loses to one agent
Team loses
Goes wrong
One agent
Parallel writes
two workers edit one module; each passes its tests, the merge fails
one writer per artifact; parallelise only the investigation
Sequential steps
worker B waits on A’s report; no parallelism, extra hops
one loop keeps the full context
Lossy briefs
“no issues found” means “could not read the gateway”
no hand-off to lose; if you split, a gaps field
Stragglers
fan-in waits for the slowest worker’s deadline
one timeline, one timeout
Nested retries
SDK × app retries: a failing leaf makes 9 calls, 27 at three layers
retry at one layer
Debugging
root says success; two children failed underneath
one transcript to read
Price it per finished task, not per call
Setup
Per task
Per finished
One agent, Opus 5.5
55K in · 5K out · $0.32 · 60% succeed
$0.53
Team: Opus 5.5 lead + 4 Sonnet 5.5 workers
205K in · 21K out (3.8×) · $0.76 · 80% succeed
$0.95
Illustrative. Lead: 45K in, 5K out = $0.28. Each worker: 40K in, 4K out = $0.12. The team is 33% more likely to finish and still 1.8× the price per finished task — worth it only when a miss costs more than the gap.
REAL SYSTEM
Illustrative example: a four-agent bug-fix team (planner, coder, tester, reviewer) passed 122 of 200 issues (61%) on its eval at 3.1× the tokens of one agent, which passed 132 (66%). Most losses were coder and tester disagreeing about an interface. The team collapsed it to one agent plus a read-only review subagent, and kept a fan-out only for “find every call site” across the monorepo — the one step that was broad, read-only and parallel.
FOLLOW-UP TRAP
“Anthropic’s research system beat a single agent by 90.2%.” — On breadth-first research, where reads split cleanly, and in the same write-up token usage alone explained 80% of the variance on BrowseComp. Much of the win is the tokens. So the honest baseline is not one agent on default settings; it is one agent at higher effort, on the same budget. Beat that before you pay for a team.
