Practitioner says multi-agent workflows work better with 1–2 active agents, then long autonomous runs

edgarpavlovsky · x · 2026-07-28

The author shares a practical framework for working with interactive multi-agent systems:

Running many parallel agents gives an initial dopamine rush, but quickly leads to burnout because the user has to manage too much context, which can devolve into AI slop.
In practice, they found it better to keep the number of heavily interacted-with agents low, usually one or two, with short runs of under 30 minutes.
The next step is to delegate much longer tasks, from more than an hour up to 8–12 hours, while parallelizing again only when the agent work feels more autonomous.
The preferred workflow becomes a cycle of planning, sending the agent off, taking a breather to reset context, and then returning to think deeply about another independent problem.
For them, preserving space for deep thinking matters more than maximizing agent count.

More from coding & agent