1,000 agents ran, only 40-60 were ever calling the model
Thomas Wolf's team scaled a swarm of DeepSeek V4.1 Flash agents past 1,000 on a couple of GPU nodes, then watched the bottleneck move off the GPUs entirely and into git merges and context re-reading.
Receipts
Thomas Wolf ran 1,000 agents and the GPUs were never the problem
Thomas Wolf's team scaled over 1,000 DeepSeek V4.1 Flash agents on a couple of B200 nodes to build a rule-based chatbot, and found the bottlenecks everywhere except the GPUs. Agents sat waiting on git merges, which forced a move to centralized merging, and on tests, and they read roughly 130 tokens for every token written.
In some of our runs, at any given moment, only 40-60 of the 1,000 agents were actually calling the model.
Willison published external sandbox prototypes for Claude Code, Codex and Pi
Simon Willison described the external sandbox pattern, where the agent runs on your Mac but executes its code in a WebAssembly, QEMU or Apple Container sandbox, with demos for Claude Code, Codex and Pi.
How they actually work
Pocock is testing harsh lint rules as a way to constrain agents
Matt Pocock's experiment of the day is an extremely constrained version of the tools he likes, enforced through what he calls incredibly harsh lint rules. The example: the agent may use React, but all state has to go in Zustand or useReducer. He is thinking of calling it /constrain.
Strip out all the nuance, create a black hole of success, all through incredibly harsh lint rules.

