A to-do list beats a manager

When Stanford gave a multi-agent team a manager, it spent 73% of its time waiting. But replace the manager with a log and to-do list? Waiting fell to 2%, and the team finished 5.5x faster.

Unlike Susan, who never reads the comments you leave her, agents don’t need managers, because they actually do their homework. (Sorry, Susan) 

But isn’t that just like an agile team with a kanban board? Sorta. Agile teams have run on a shared to-do list since 2001. But, that’s like a grocery list on the fridge. Everyone writes on it. Nobody reads it. And the house ends up with three cartons of milk and no eggs.

In 2006, Microsoft found its own engineers avoid written records of what they already know. Fair enough. The documents sat on internal sites with no working search, and about half were out of date. Reading them was like trusting the shopping list from last March.

But an agent reads everything. In 80% of runs, an agent went back for a teammate’s updated work. When one agent claimed a job, another agent already doing that job read the claim and dropped it. 

Humans would only have found out at the sprint review.
Then would’ve booked an “alignment” meeting about it. 

But at least a manager catches problems, right? Nope. In another setup, a helper agent found that a speed shortcut gave wrong results in some cases, the manager decided the speed was worth it, and the run failed its tests. On the shared to-do list, a warning stays open until someone fixes it.

So human teams need managers…but for a duller reason than we think: Someone has to read everything so nobody else has to. Agents do their own reading. That makes this a new kind of organization, one agile drew 25 years ago and nobody could reliably staff.

This post is 327 words about how nobody reads.
If anyone finished it, it was probably an agent.

—

Sources:
“Decentralized Multi-Agent Systems with Shared Context,” Mao et al, Oct 2026
“Maintaining Mental Models: A Study of Developer Work Habits,”LaToza et al, May 2006