A client asks for an AI team. Not an automation, a team: a researcher agent, a writer agent, a reviewer agent, a manager agent to boss them around. It sounds impressive on a slide. It sounds like billable scope, too.
Before you sell it, read what The Decoder reported from Vals AI's testing. Across four comparisons of agent teams against solo agents, only one showed a statistically significant quality gain, and the teams cost between 1.8x and 5.1x more. The results suggest the extra spend isn't worth it in most cases.
That is a coding benchmark, not your client's onboarding workflow. Don't over-read it. But it points at something you already suspect if you have ever managed people.
Why this lands on your margin
If you run delivery for clients, architecture is not a technical detail. It decides your cost of goods, your support load and how often you get paged.
An agent team that costs several times more to run for output that is rarely better creates a bill with only three places to go:
- You absorb it, and your margin on the retainer shrinks quietly.
- You pass it on, and the client eventually asks why it costs so much.
- You hide it in a flat fee, and it bites when usage grows.
None of those is a good place for a founder who is already the bottleneck. More moving parts also means more things to debug at 9pm, and you are probably still the one debugging.
The reporting also cites OpenAI researcher Noam Brown saying multi-agent systems mainly buy speed, not quality, and that some work parallelizes well while other work does not. His comparison was that throwing 10,000 agents at a novel is as pointless as throwing 10,000 people at it. Anyone who has added a fourth person to a late project knows the feeling. Coordination eats the gain.
The default: one agent plus a check step
Here is the position I would take with any client: start with one well-scoped agent and a checker, and make every additional agent earn its seat.
The checker matters more than the headcount. It can be a second pass that compares the output against the brief, a set of rules the output must meet, or a human approval at the one point where mistakes are expensive. This is the same logic you use with staff: one capable person with a reviewer beats five people passing work around with no one accountable.
A single agent with a check step gives you:
- One place to look when something goes wrong.
- A predictable cost per run, which you can price into the engagement.
- A clear failure signal, because the checker either passes the work or it does not.
A decision test before you add a second agent
Adding agents is not forbidden. It just needs evidence. Run this test before you quote a team.
1. Is the work actually parallel?
Research across many independent sources can split cleanly. A single deliverable that has to read as one voice usually cannot. If the pieces depend on each other, more agents mostly means more handoffs.
2. Is speed the thing the client is paying for?
Per the reporting, teams tended to buy speed rather than quality. If the job runs overnight and nobody is waiting, speed has no value, and you are paying extra for it.
3. Did the team beat the solo version on your own samples?
Build both. Run each against the same 20 or so real examples from the client's work. Score them against the same checklist. If the team does not win by a margin the client would notice, ship the solo version.
4. What does the gap cost per run?
Multiply the extra cost by the monthly volume. Then ask whether the quality gain, if any, is worth that number to the client. This is first-principles arithmetic, and you should do it with the client's real volumes, not a guess.
If you want a method for the scoring step, our piece on building a production-grade AI workflow with real evals covers how to turn a client's idea of good into something you can test.
A worked example (illustrative, not a guarantee)
Suppose a client wants weekly competitor summaries. The team design uses four agents: a collector, an analyst, a writer and an editor. The solo design uses one agent that does the whole pass, followed by a checker that verifies each claim has a source in the collected material.
Say the team costs three times as much per run. That multiple is invented for this example, so treat it as illustrative, not a guarantee. If your side-by-side test shows the summaries are about equally good, the extra two-thirds of the bill buys nothing the client can see. If the team wins clearly on one stage, say collection from many sources, you can split only that stage into parallel workers and keep the rest solo.
That is the pattern worth aiming for: a mostly simple system with agents added at specific points where a test justified them.
The honest tradeoffs
The evidence has limits, and you should say so to clients.
The Vals AI result came from one coding benchmark and a handful of comparisons. In that same reporting, one of the four did show a real gain, and Anthropic's own tests showed larger teams reaching a performance level faster. Other tasks, models and setups may behave differently, and these tools change quickly. A single agent will also hit a ceiling on some jobs, particularly long ones with many separate pieces.
There is a second tradeoff on the sales side. A team-of-agents diagram is easier to sell than a single agent and a checklist. If you quote the simpler design, some clients will think they are getting less. The answer is to show them the test: here is the solo version, here is the team version, here is what each costs and how each scored. Clients who respect evidence will respect that. Clients who only want the org chart may not be a fit anyway.
Finally, simple does not mean careless. A single agent with no checker is just a faster way to ship mistakes. The check step is the part you cannot skip.
What to do this week
Pick one live or proposed client build that uses several agents. Build the single-agent-plus-checker version next to it, run both on real samples, and write down cost per run and quality score for each. You will learn more from that one comparison than from any benchmark, including this one.
If you would like a second set of eyes on how to scope a build like that, or on where your own delivery architecture is costing you margin, the free 30-minute AI Opportunity Audit is a no-pressure place to start. You keep everything we map out, whether or not we work together.
Visuals in this article are AI-generated.
Sir Tay Jackson
Founder, Air Flow Automations
Founder of Air Flow Automations. Builds custom AI systems for agency owners, e-commerce founders, and service operators who want their time back without adding headcount.
