KorvaJoin the waitlist
Tech

Stop Giving AI Agents Job Titles and Hire a Temporary Team Instead

A lead agent staffs each job with temporary workers, and Anthropic found the approach scored about 90% higher than using one agent alone in its tests.

AI image
The AI Daily Brief: Artificial Intelligence News and Analysis · 7 Things AI-Native Companies Do Differently

Early this year, Nathaniel Whittemore says he had built himself a small digital staff. He had research agents, writing agents, project coordinators, and a chief-of-staff agent sitting above them all.

Now, he argues, that setup was a mistake. "Agents as teammate is in fact the wrong mental model," he says.

The teammate idea helped people see AI as something that could go do work on its own, not just answer questions in a chat window. But Whittemore says a fixed roster, with permanent names and job titles, does not match how sophisticated users now work.

Instead, he says, one lead agent acts as an orchestrator and hires temporary subagents for the task. Describing Anthropic's research system, Whittemore says it spins up "One agent for a simple fact check, 10 or more for a hard problem." In Anthropic's own research tests, that setup scored about 90% higher than using one agent alone. Each worker gets narrow instructions and separate context for its piece.

Whittemore argues that fixed roles create costly handoffs. He cites Anthropic's finding that splitting agents into planner, coder, tester, and reviewer tends to backfire, with agents spending more effort coordinating than working, like a game of telephone.

He has made the same change himself. He now keeps persistent project context in shared repositories and spins up agents to do specific jobs inside them. The software company Every, he notes, similarly abandoned one proxy agent per person for just a few shared agents able to expand through subagents as needed.

His preferred analogy is an on-demand consulting firm. A goal triggers an engagement, the lead staffs invisible workers, the client gets an evidence-backed deliverable and owns the result, and the team disbands when the work ends while the records, and the method, remain.

Questions people ask

What does the lead, or orchestrator, agent actually do?

It plans the job, breaks it into pieces, and hires subagents to do each piece. It decides how many workers are needed, which model each should use, and what information each gets, then collects their outputs into a final result.

If the agents disappear, what stays permanent?

The shared context stays. That includes documents, decisions, project history, and instructions that future agents can reuse. In Whittemore's setup, that lives in repositories, so any new agent can drop in and get up to speed.

Who is accountable when a temporary team does the work?

You do. In the consulting model Whittemore describes, you bring the goal and get back a deliverable with evidence. You sign off and own the result internally.

More in Tech