Agent swarms are crossing into genuinely strange territory
A thread on 'Agent swarms and the new model economics' describes a system hitting 1,000 commits per second, orders of magnitude beyond what any human team produces. To support that throughput, the builders had to rethink version control primitives entirely. The example task used to benchmark the swarm: building SQLite from scratch in Rust, using only the documentation as input. That is not a toy benchmark.
A separate but adjacent thread argues that you 'only need the frontier model for one single edit,' with the implication that most of the actual work in an agentic workflow should be done by cheaper, faster, smaller models. Commenters are already experimenting with phased implementation plans where a main agent delegates to focused subagents per phase. These two threads are pointing at the same thing: the architecture of AI-assisted software development is being rebuilt from scratch right now.
The counterpoint in the comments is healthy skepticism about cost. Running a swarm at 1,000 commits per second is not cheap, and the economics only work if the output quality is high enough to justify it. The thread acknowledges this is still experimental, but the framing is 'glimpse into the future,' not 'interesting toy.'
So what?
The implication for founders building developer tools or internal platforms is that the unit of work is shifting from 'what can one engineer do per day' to 'what can one engineer orchestrate per minute.' If you're building tooling, your abstractions need to handle parallelism and coordination at a scale most existing tools were not designed for. And if you're running a software team, the question of how to structure agentic delegation is not a future problem.