A team of agents: AI Engineering, story 16
One agent splits a research job between three helpers. See what that buys, what it costs, and what gets lost on the way.
When one agent hands work to others, what that buys, and what it costs.
The story, step by step
A job with three parts
HM Nayem wants advice for the hiking club: which of three GPS watches to buy. Each watch means reading a dozen reviews and spec pages. Three similar pieces of research, none depending on the others.
One agent could do it alone
A single agent can do this the way the coding agent worked: one watch after another, round by round. But it would be slow, and all 45 web pages would pile into one window. You have seen what a crowded window does.
Split the job
So the lead agent does what it did with a to-do list, with one difference. Instead of working through the items itself, it will hand each one to a helper. Each helper is a full agent of its own: a model, tools and a loop.
A helper starts with nothing
This is the part people miss. A helper begins with a completely empty window. It has not seen HM Nayem's message or the lead's thinking. Whatever is not written in its brief does not exist for it. Writing a good brief is the whole skill.
All three at once
The other two briefs go out, and all three helpers start searching at the same moment. Nothing makes one wait for another, because the three jobs do not depend on each other.
Each fills its own window
Every helper reads about fifteen pages, and all of that lands in its own window, not the lead's. The mess of research stays where it was made.
Only a summary comes back
When a helper finishes, it does not send back everything it read. It sends three lines. Forty-eight thousand tokens of reading become about sixty tokens of findings.
The lead's window stays clear
Look at the lead's window after all that work. The job, three briefs, three short reports. Almost empty. It can now think about the comparison with nothing in the way.
The answer
The lead compares the three reports and writes one recommendation. HM Nayem sees a single agent and a single answer. The team behind it is invisible.
What the team bought
Two real gains. The work finished in a third of the time, because three helpers ran side by side. And the agent making the final call did it with a clear head.
What the team cost
Now the bill. Four agents each run their own loop, and each loop resends its own growing record every round. A team typically uses several times the tokens one agent would. Speed and clarity were bought, not found.
What gets lost on the way
There is a quieter cost. The club hikes where phones have no signal, and HM Nayem mentioned it weeks ago. The lead never put that in the briefs. So Helper 2 called the TrailPro's phone-based maps “maps”. A helper cannot ask HM Nayem what was meant.
When a team is worth it
So a team is a trade, not an upgrade. It pays when the parts are independent and each needs a lot of reading. For small jobs, or ones where every step depends on the last, one agent is cheaper and makes fewer mistakes.
Now at real speed
One job in, three briefs out, three helpers reading at once, three short reports back, one answer. More agents buy speed and clear windows. They are paid for in tokens and in details lost between them.
A team of agents is a trade, not an upgrade.
- A lead agent can hand parts of a job to helpers that run at the same time.
- Each helper has its own empty window and knows only its brief.
- Helpers send back short summaries, so the lead's window stays clear.
- A team uses several times more tokens than one agent.
- Anything left out of a brief is lost, so teams suit independent parts.
Ideas in this story
- Teams of agents: A lead agent hands parts of a job to helper agents, each with its own fresh window.
- Context engineering: Deciding what goes into the model's window at each step, and what stays out.