Sooner or later, someone will offer to sell you “a team of AI agents.” One to read the email, one to check the schedule, one to write the replies, one to check the others. It sounds like hiring a whole office for the price of software.
Before you buy a team, it is worth knowing what a team actually gets you, and what it costs.
Bigger teams are now easy to build
On October 9, Anthropic added a beta feature to its developer platform that lets one AI agent plan a job and hand pieces of it to many other agents at once. Its documentation says a single run can start up to 1,000 agents, with up to 64 working at the same time.
That is useful for some work, like reviewing hundreds of documents. But “possible” is not the same as “better,” and more agents means a bigger bill.
What a team of agents actually bought in one test
Vals AI, a company that tests AI models, ran a fair comparison. It had two leading AI models each build the same 50 small web apps, once as a single agent and once as a team. Its finding: the teams cost 1.8 to 5.1 times as much, and in only one of the four comparisons was the team’s result measurably better. In some cases, simply letting the single agent think harder helped more than adding teammates.
Vals is careful about its own limits: each setup ran once, on one kind of task. But the direction matches what owners find in practice. A team is only worth it when the work genuinely splits into separate pieces that do not need to talk to each other along the way. Vals puts it the same way: teams should gain the most when the work splits into parts that can proceed side by side.
Most office work does not split like that. Answering a customer email means reading it, checking the schedule and writing a reply, in that order. Three agents passing it along add hand-offs, not speed.
A made-up example
Here is a made-up example. A cleaning company is offered a “four-agent system” to handle new booking requests: one agent reads the request, one checks availability, one prices the job and one writes the reply.
In the first month it works, mostly. But when a reply quotes the wrong price, nobody can easily tell which agent got it wrong, and the monthly bill is several times what was expected.
The fix is one assistant with a clear job: read the request, check the calendar, draft the reply with the price from the price list, and send anything unusual to the office. One set of instructions, one place to look when something goes wrong, and one small bill.
Three questions before you pay for more than one
- Does the work truly split? Can it be broken into pieces that do not depend on each other, like checking two hundred separate invoices? If each step needs the last one’s answer, one assistant is the right size.
- Has one assistant been given a fair try? Clear instructions, the right access, a narrow job and a person checking the unsure cases. Most problems are fixed there, not by adding agents.
- Do you know what a good result is worth? If a team costs several times more, you need a number for what the better result saves you. Measure cost per finished job, not per message.
If any answer is no, start with one.
How to test it fairly
If a vendor says a team will do better, ask for a fair test. Run both on the same real work from your business, more than once, because a single good run can be luck. Count the cost of every finished job, including the retries. And have a person spot-check a sample of the results rather than trusting either system’s own report.
Your one step this week: if you use or are being sold an AI system with several agents, write down in one sentence what each agent does. If two sentences describe one job, you are probably paying for a hand-off you do not need.
If you want a second opinion on what to automate and at what size, here is what we build.
Sources
- Claude Platform release notes, October 9, 2026 — Anthropic
- Workflow runs: limits — Anthropic documentation
- Do Agent Teams Pay Off? A Case Study on Vibe Code Bench — Vals AI
Kush AI Automation builds in accounts you own, and our monitoring alerts when a run fails and when it never starts, so an automation that quietly stops still reaches a person. Bring the job you are thinking of handing over and book a free call.
