An agent team: the trip HQ
Lesson 15 of 22 in the course “AI Agents for Life”. Free, with animations and a quiz.
One agent or a team
Liza and three friends are going to Lisbon for 4 days. Hand it all to one agent and it starts with tickets, drowns in accommodation prices and by the end forgets the budget was €400 per person. Remember from the memory lesson: an agent’s context window is limited. The way out is to split the work like a normal team. The lead breaks the task down and hands out the parts. Three helpers each search for their own piece without getting in each other’s way. Then the lead puts it all into one plan and checks the budget.
It works: Anthropic’s example
That’s how Claude’s research mode works. A lead agent makes a plan and launches several helpers in parallel; each one searches its own part and brings back a summary. In June 2025 Anthropic published a breakdown: on their internal research tasks, this kind of team did 90.2% better than their strongest model of the time, Claude Opus 4, working alone. Why? The helpers search at the same time, and each has a clean context window for its own part. An important caveat: these are information-search tasks where the parts are easy to split.
The cost of a team
There’s no such thing as a free team. According to the same Anthropic breakdown, one agent uses about 4 times as many tokens as a normal chat, and an agent team about 15 times as many. Tokens are what your limit and your bill are made of. So the rule is simple: a team is for a big task where the result is worth the money. For “what time does the museum open” one chat is enough. Anthropic says the same: simple fact-finding needs just one agent and a few searches.
When a team, when not
A team is strong when the task can be cut into independent pieces. A trip: tickets, accommodation and places are searched separately. Choosing a university: each helper gathers data on its own two universities. But one essay can’t be cut up: the third paragraph depends on the first, and the helpers would write three different texts. Anthropic says plainly that teams are still a poor fit where everyone needs one shared context and there are many dependencies, as in most coding tasks. The test is simple: could you hand the parts to three people without them calling each other every five minutes?
How the lead hands out tasks
The most common way a team breaks is vague assignments. The same Anthropic breakdown has an example: on a question about the chip shortage, one helper studied the 2021 crisis while two others duplicated each other and dug into the same thing. The engineers’ takeaway: each helper needs a goal, an output format, guidance on where to look and clear task boundaries. In other words, a mini-brief like in level one. “Find a place to stay” is bad. “Find 3 central places for 4 nights for 4 people at up to €60 per person per night, as a table, don’t look at tickets” is good.
Try it: the trip HQ
You can assemble a team by hand, with no services at all. Step 1: paste the lead’s brief below into a chat and fill in your trip. It returns three assignments. Step 2: open three separate chats or three deep research runs and give each its assignment. Step 3: paste the three answers back to the lead, and it puts together the plan and checks the budget. It’s slower than one button, but you see how a team works from the inside, and nothing gets booked without you anyway.
Check yourself
- Which task needs an agent team?
Correct answer: Compare 6 universities on cost, housing, rankings and scholarships. Six universities are easy to hand out two per helper, and the parts don’t depend on each other. That’s exactly where a team wins. - Why did that happen?
Correct answer: A team uses about 15 times the tokens of a chat. Each helper searches and reads on its own, and the lead also assembles. Hence the cost: for simple questions it’s total overkill. - What should the lead do differently?
Correct answer: Give each one clear boundaries: “only accommodation, leave tickets alone”. It’s the same mistake as in Anthropic’s chip example. Clear boundaries in each assignment remove duplicates. - What should she do with this plan?
Correct answer: Re-check the prices on the websites and only then book herself. Hotel prices change daily, and the helpers may have picked up old ones. The brief itself marks what to re-check, and the money stays under your control.