Boris Cherny describes a key insight behind agent topologies: “Multiple agents, they have fresh context windows that aren’t essentially polluted with each other’s context or their own previous context. And if you throw more context at a problem, that’s like a form of test time compute.”
The practical proof: Claude Code’s plugins feature “was entirely built by a swarm over a weekend. It just ran for a few days. There wasn’t really human intervention.” An engineer gave Claude a spec and told it to use an Asana board. Claude created tickets, spawned agents, and the agents started picking up tasks independently — each with its own uncorrelated context window.
Boris also reveals that “the majority of agents are actually prompted by Claude today in the form of sub agents. A sub agent is just a recursive Claude Code” — internally called “mama Claude.” His personal workflow mirrors this: “80% of my sessions I start in plan mode. Claude will start making a plan. I’ll move on to my second terminal tab and then I’ll have it make another plan. When I run out of tabs I open the desktop app.”
This extends Treat AI like a distributed team, not a single assistant with a theoretical foundation — it’s not just about parallelism for throughput, it’s that independent context windows produce qualitatively better reasoning by avoiding cross-contamination. It also connects to The context window is the fundamental constraint — everything else follows: if context is the bottleneck, the solution isn’t just managing it better within one window — it’s multiplying windows.