Shumer says one manager agent is enough, and his mom's closed $600
The founder says he runs hundreds of agents a day and still only needs one session that farms work out to the rest.
Matt Shumer says the elaborate multi-agent setups going around are fun and not actually productive. His argument is that one agent session acting as a manager, spawning others as needed, is all you need, and that if something is off you tell the main session and it adjusts the approach. He says this comes from running hundreds of agents a day, dozens at a time, with spikes into the hundreds at once.
Then he handed that setup to his mother. She named it Morgan, and Shumer reports it closed a $600 ad deal for her site. His read on why it worked is the Bitter Lesson applied to agents, ask the model and let it spawn what it needs.
She named him Morgan. He just closed a $600 ad deal for her site.
@levelsio landed in the same place from a blunter angle, calling software factories idiocy and saying one AI coding agent is good enough because it can spin up more agents if it needs to.
The other side of the argument
Greg Isenberg spent 31 minutes defending the idea, if not the name. His version of a software factory is five or six markdown files sitting next to your code telling agents how you like to work. Each feature gets its own branch so multiple agents work without stepping on each other. A build station sets code rules, a proof station makes the agent produce evidence such as screenshots, videos and speed numbers, and a review agent sends anything that fails back through the line. You show up at the end to merge.
His case is about the bottleneck. Without structure you build one feature at a time while prompting, checking diffs and testing yourself. With it, he says, ten or more features run in isolated copies and you approve finished work that already has evidence attached instead of supervising it.
The week had plenty of people in the middle. Hiten Shah ranked the harnesses he uses outside the big labs, putting Hermes Agent first because it let him put multiple agents to work fastest without assembling the system himself. Arvid Kahl asked, plainly, how anyone avoids losing their mind using worktrees in Claude Code for parallel agentic coding.
The business read
Both camps are made of people running agents at volume, and they disagree about where your effort goes, into the model and the prompt or into the structure around the output. Guillermo Rauch pointed at the thing underneath both positions, that agents are only as good as the proof-checkers, compilers, type systems and linters you give them. The gate matters more than the orchestration. Nobody in the sources has shown numbers comparing the two approaches, so pick the one that matches how much you trust your own review.
Agents are only as good as the proof-checkers, compilers, type systems, and linters you give them.
Just using one AI coding agent is good enough and it can spin up more agents if it needs to