There is no 'AI' in 'Team'
Scott Werner, who is frustratingly good at writing what I'm thinking about LLMs, has a new post out where he compares being an "agentic" coder to being an octopus, with each arm being a separate instance of Claude Code independently thinking and acting on its own. It's a good post and you should read it.
In the middle, he said the thing that was what first came to mind when I saw the image of the octopus in this context:
Here's the thing about teams now:
Two developers on one codebase is like two octopuses sharing one coral reef. Technically possible. Practically ridiculous. Everybody's arms getting tangled. Ink everywhere. The coral is screaming (coral doesn't scream, but work with me here).
But one octopus with eight projects? That's just nature.
The more time I spend with coding agents, the more I become convinced that they are damn-near incompatible with working in teams. I've suggested this before, but I really think more people should be chewing on this. The bottleneck for software teams—the thing that's always made them less than the sum of their parts—is the handshake problem. It's the one thing from The Mythical Man-Month everyone remembers: "Adding manpower to a late software project makes it later."
For the last 50 years, this has been (quite reasonably) understood as the number of humans on a team. That the number of relationships between those humans in an organization's design could be used to compute an approximate productivity tax on the collective's broader efforts to encode some kind of intention into software.
If we have 8 humans working on a software project, we have (8 ⨉ 7) ÷ 2
or 28 relationships to manage, with each individual capable of and burdened by the need to bidirectionally seek shared understanding and consensus with one other for the purpose of coordinating their efforts.
Now imagine the team having 8 humans each juggling 8 sub-agents on a project. This figure balloons to (64 ⨉ 63) ÷ 2
or 2,016 relationships. In the good ol' days of 2022, this quadratic increase in communication cost for squaring the size of the team was enough to give people pause on its own. But when ⅞ of the team are AI agents, it adds an all-new wrinkle to the math: 1,764 of those connections are unidirectional. The agents can receive information and instruction but they lack durable institutional memory, they can't pipe up in meetings, and each has its view of the world bottlenecked behind a single operator. Each of those complications has the opportunity to dramatically compound the already-really-quite-bad errors we typically associate with large software teams. This is made even worse by the fact that a manager has no observable signal that their team's composition has changed so radically—they'll walk into the room and see the same eight nerds staring at their computers as ever before.
My theory on why this issue hasn't already triggered productivity meltdowns is a happy accident of circumstance, owing to the fact that the people currently trailblazing multi-agent workflows in earnest are highly-engaged, driven programmers—the go-getting early adopters who were using Rails in 2005 and Node.js in 2009. As a result, the median team of eight engineers may not even have one such developer—which means we haven't seen what deploying coding agents at scale will really look like yet. My prediction is that as these tools continue to go mainstream, things are going to go about as well as if you were to throw a dozen octopuses in an aquarium together.
If none of this is quite clicking with you, think of it this way. Team A has 8 programmers in a room working on a project. Team B has 8 technical analysts each managing a separate sub-team of 8 offshore developers somewhere in South Asia, replete with all the time zone and communication constraints those impose. We have a lot of data to indicate that Team B is in for a bad fucking time, but that scenario is effectively what mainstream adoption of coding agents as they exist today would represent.
Anyway, if your entire team are working to keep their coding agents' hoppers full (replete with subagents and juggling multiple tasks at a time), what is your effective team size by this measure? Am I wrong here? Is everything actually going great? Is coordination not suddenly much harder than it was before? Let me know.