Language models can now learn from thousands of parallel experiments without losing accuracy

A new system called Combee lets AI agents run in parallel while learning from all their combined attempts at once, instead of learning from one attempt at a time. This means AI systems can improve themselves faster and cheaper by scaling up the learning process without the usual quality loss that happens when you run too many things at once.

AI agents already learn from their own mistakes — they run a task, see what went wrong, and adjust their instructions for the next attempt. But that learning has been slow because it happens serially, one experiment after another. Combee removes that bottleneck by letting hundreds or thousands of agents run in parallel while the system learns from all of them together without degradation. What becomes possible: agents that improve themselves in hours instead of days, and companies can run larger-scale experiments without the math falling apart. What stays the same: the fundamental limitation is still human-defined tasks — this just lets you iterate through more of them faster.