Article Takeaways / TL;DR
- The “AI employees” you’re seeing everywhere borrow their shape from an org chart. It looks like progress, but every handoff is a place for context to thin out.
- The engineering guidance underneath is more restrained: start simple, and add agents only when the work actually splits that way. Anthropic’s multi-agent research system won on open-ended research, at roughly fifteen times the tokens of an ordinary chat.
- Most business work doesn’t split like that. It needs one clear owner, and another agent inside it is a cost, not a given.
- Before adding the next agent, name what it improves and whether that beats the handoffs, tokens, and upkeep. If the only gain is a better-looking org chart, you’ve built the demo, not the system.
Open any feed right now and someone is introducing their AI employees. Brad runs sales, Harper writes the posts, and an intern named Skye never sleeps and works the DMs. It gets sold as a staff you never have to pay. A lot of it is a tiny bureaucracy that burns tokens to argue with itself.
The pitch borrows its shape from an org chart, which is why it feels right. Give each bot a job title and the whole thing looks like a company, and a company looks like progress. What it also adds is coordination overhead. Every handoff is a chance for context to thin out, and a badly built one turns your specialists into a game of telephone, each bot working from a slightly worse copy of what the last one knew.
The engineering guidance underneath all this is far less theatrical than the sales pitch. Anthropic’s advice to developers is to start with the simplest thing that works and add complexity only when the task justifies it. Sometimes that means not building an agent at all.
That is not an argument against multiple agents, and the honest version is more useful than the hype. Anthropic also built a multi-agent system for Claude’s research feature, and it beat a single agent handily on open-ended research. The token overhead is in the same write-up: in their data, multi-agent runs used roughly fifteen times the tokens of an ordinary chat, against about four times for a single agent.
The approach pays off when the work fans out into independent threads that run at once. In that 2025 write-up, they noted coding offered fewer clean splits than research. The takeaway was never “more agents” or “fewer agents.” It was that the architecture has to follow the shape of the task.
That is the principle we’re building our own content engine around. It will draft scripts and schedule social video, and it would be easy to give each stage a named agent and call it a crew. We’re designing it as one workflow with one owner. In the research step, separate searches run independently and come back in a shared format. Scripting and production stay sequential, because each stage carries forward decisions made in the one before it. The payoff we’re after is a system we can reason about, not a crew we have to referee.
Tool overload is becoming a weaker reason to split work across agents too. A single agent used to choke on its own tools, since loading dozens of verbose definitions upfront eats a real share of the context window before any work starts. Anthropic’s Tool Search, introduced in late 2025, lets one agent discover the tools it needs instead of carrying every definition. A large toolset no longer automatically calls for a larger cast.
This is worth caring about even if you will never configure an agent, because you inherit what the diagram leaves behind. The handoffs and the upkeep land on you after the demo is over. Someone still has to own the outcome, and every agent inside it is a cost you take on, one that has to return more than it spends.
So the question is not how many AI employees you can spin up. Before you add the next one, name what it improves, and how you will know whether that improvement is worth the handoffs, the tokens, and the upkeep. If the only thing it improves is the org chart, you’ve built the demo, not the system.
Common questions
Are multi-agent setups always a bad idea?
No. They earn their place when the work genuinely splits into independent parts that can run at once, like broad research. Anthropic built one for exactly that. The cost is real, though, roughly fifteen times the tokens of a chat, so the task has to be worth it.
How do I know whether my workflow needs one agent or several?
Ask whether the work comes apart into pieces that don’t depend on each other. If it does, parallel agents can help. If each step builds on the last, which is true of most content and build work, one agent with a clear playbook holds it better.
Isn’t a single agent limited by how many tools it can handle?
That used to be the main reason to split work across agents. Newer tooling lets one agent pull only the tools it needs for a task instead of loading all of them at once, so a large toolset no longer forces a larger cast.
What should I actually ask before adding another AI agent?
Name the job it does that your workflow can’t already cover, and be honest about what the extra handoffs, tokens, and upkeep cost to run. If the only thing it improves is the diagram, it isn’t earning its place.
Author:
Jeremy Bokor
Founder, Nifty Inc
