106 Agents. $76. One Question.

I asked a deep-research harness one hard infrastructure question. Without my realising the scale, it fanned out into 106 separate agents, processed about 32.5 million tokens, ran for 33 minutes and cost roughly $76.

The same question asked of a one-dollar research tool would have taken about a minute.

So was the expensive version 76 times better? No. But it did one thing the cheap version structurally cannot, and there is a real question underneath about whether a fan-out like this is engineering or consumption dressed as rigour. Here is the breakdown I wish I had before I hit go.

The question

It was a genuinely hard architecture question, the kind with no clean answer in any documentation:

How do you let thousands of customers point their own custom domains at a self-hosted, multi-tenant platform, without handing the whole problem to a managed service, and what breaks at scale?

Certificate issuance limits, multi-tenant database design, cross-server topology. The sort of decision where a wrong assumption costs months, and exactly the sort you want verified rather than vibed.

So I reached for the heavy option: an agentic workflow that fans out searches, fetches sources, adversarially fact-checks every claim and synthesises a cited report.

I did not appreciate what "fans out" meant.

Anatomy of 106 agents

The harness does not make one large call. It decomposes the job into a pipeline, and each stage spawns its own fleet:

11 Scope agent          decompose the question into 5 angles
2       │
3       ▼
45 Search agents        one per angle
5       │
6       ▼
724 Fetch agents        read sources, extract claims
8       │
9       ▼
1075 Verify agents       3 skeptics per claim
11       │
12       ▼
131 Synthesise agent     merge and cite

The verify stage is where the population lives. Every factual claim pulled from the sources goes to three independent skeptics, each instructed to try to refute it. A claim survives only if they fail. With 25 claims under test, that is 75 agents whose entire job is fact-checking:

Seventy-one percent of the run was checking, not finding. That is the design working as intended, and it is also the bill.

Where the $76 went

Across all 106 agents: about 32.5 million tokens, roughly $76 at frontier-model pricing. The surprising part is which line dominated.

1Cache WRITE   $46.35  ████████████████████████████████████████  61%
2Input         $12.88  ███████████                               17%
3Cache READ    $11.15  ██████████                                15%
4Output         $5.74  █████                                      8%
5                     ─────────
6Total                 ~$76.12

The counter-intuitive lesson is that the expensive part of multi-agent research is not the thinking. Output was 8% of the bill. Sixty-one percent was the setup tax of spawning a hundred short-lived agents, each writing its context to cache and none living long enough to amortise it.

More agents means more cache-write tax, largely regardless of how much work any one of them does. That is the opposite of how caching behaves in a long-running conversation, where the write is paid once and read back for the rest of the session. Wide and shallow is the shape that pays the write over and over.

One note on how you pay. On a metered API key this is a literal charge. On a subscription you see no dollar figure at all, just your usage allowance draining about 76 times faster than a normal query. Same dynamic, different meter, and the second one is much easier not to notice.

Is this useful, or is it burning tokens?

Worth saying out loud: a tool that profits from token consumption just spun up 106 agents where one might have done. That deserves scrutiny rather than a shrug.

The case that it is real work. It was opt-in; the default for a normal question is one agent, and I typed the command that chose otherwise. The agents did distinct work: I pulled the transcripts, and each verify agent independently re-read a claim against its source and voted. And output was 8% of the cost, which is not what padding looks like. Something engineered to inflate usage would produce bloated generation, not spend on reading and checking.

The case for keeping a hand on your wallet. It did not tell me the scale or the cost before running. There was no "this will spawn about 100 agents, cost about $76 and take 33 minutes, proceed?" I found out by inspecting the run afterwards. Opaque defaults that happen to consume a lot are exactly where a metered vendor's incentives and your budget quietly diverge. The synthesis was also thin for the price, and the fan-out scales cost faster than it scales assurance.

The honest verdict is that it is a genuine power tool whose economics favour the vendor when you do not right-size it. The fix is not distrust, it is deliberate escalation and demanding that tools surface cost up front.

The comparison nobody wants to run

Read that table honestly. On breadth, sourcing and readable synthesis, the cheap tool is competitive or better, at one to two percent of the cost and three percent of the time. The harness fetched the same tier of blog content, and its final summary was thin enough that I wrote the actual architecture recommendation by hand.

Most of what I paid for, I could have had for about a dollar. Except for one column.

What the extra $75 actually bought

The verify stage killed four confident-sounding claims that a single-pass tool would have handed me as fact:

These are precisely the plausible half-truths that a research tool launders into facts, and the reason retrieval quality is a question about the query rather than the index. The skeptics caught them, and just as usefully they confirmed the load-bearing constraints from primary sources: the certificate authority's real rate limits, the sanctioned and unsanctioned ways around them, and the exact software version that fixed a known multi-tenancy bug.

That is the actual product. Not better research, verified research. About three dollars per fact-checked claim to know which facts are real.

For a blog post that is absurd. For a decision you are about to build a platform on, catching four wrong assumptions beforehand is cheap.

The tiering I follow now

1COST          TOOL                        USE WHEN
2─────────────────────────────────────────────────────────────────────
3~$0.01-$1     Single AI answer or         Default. Facts, current
4              cheap research mode         events, "what's the state
5                                          of X", first-pass scans
6─────────────────────────────────────────────────────────────────────
7<$1           One scoped search agent     Searching a large codebase
8                                          or corpus broadly
9─────────────────────────────────────────────────────────────────────
10~$10-15       Capped multi-agent          You want verification but
11              (fewer angles,              not the full bill
12              single-vote check)
13─────────────────────────────────────────────────────────────────────
14~$50-80       Full adversarial harness    Rare. High-stakes, build-on-it
15              (100+ agents, 3-vote)       decisions where a wrong
16                                          assumption is expensive, and
17                                          only when chosen deliberately
18─────────────────────────────────────────────────────────────────────

The economics in one line:

1Value of multi-agent research  ≈  (cost of being wrong) × (claims you
2                                   can't afford to take on faith)
3                                  ────────────────────────────────────
4                                     ~$3 per verified claim

If being wrong means looking slightly silly in a blog post, skip it. If it means rebuilding the platform next quarter, it pays for itself on the first killed claim.

Six things worth keeping

1. Deep research fans out further than you think. It can mean a hundred agents, not one. Know the scale before you start.
2. The cost is in the fan-out, not the thinking. Output was 8% of the bill and cache writes were 61%. Price scales with agent count more than with depth.
3. For breadth and a readable answer, cheap tools are competitive. They matched the harness on sourcing and beat it on synthesis, for about 1% of the cost.
4. What you pay a premium for is verification. Adversarial fact-checking is the thing a single-pass tool cannot do, and it earns its keep exactly when a confident wrong answer is expensive.
5. Mind the incentives without crying foul. On inspection this did real work rather than padding. It also did not surface the cost up front, and that is the part to stay alert to.
6. Default to cheap, escalate deliberately. Make the expensive tool an eyes-open choice, never a reflex.

The tools are remarkable. The trick is matching the weight of the tool to the stakes of the question, and not letting a one-word request quietly become a hundred-agent run.

Numbers here are from a single real run on 2026-06-10: 106 agents, about 32.5M tokens, 33 minutes, roughly $76 at the frontier-model rates of the time. Token counts came from the run's own agent transcripts. Yours will vary with question complexity and pricing.

FAQ

Why does multi-agent research cost so much more than a single query?
Not the thinking. Output was 8% of the bill; 61% was cache writes from spawning a hundred short-lived agents that never amortised them.

Is a hundred-agent research run actually better than a cheap one?
On breadth, sourcing and synthesis, no. On verification, yes, and that is the only column where it wins.

What is adversarial verification in AI research?
Each claim goes to several independent agents instructed to refute it. It survives only if they fail. Here that was 75 of the 106 agents.

When is the expensive option the right call?
When being wrong is expensive and there are many claims you cannot take on faith. At roughly $3 per verified claim, that is a platform decision, not a blog post.