AI agents are getting better at learning from their own history, and a new framework called Combee removes a major bottleneck in that process.
Combee tackles what researchers call prompt learning: instead of retraining a model's weights, the system rewrites the instructions an AI agent uses based on lessons from earlier runs. Existing tools like ACE and GEPA do this well for a single agent or a small number running in parallel, but they lose accuracy once you try to scale up to many simultaneous agents. Combee is designed to absorb lessons from many agentic traces at once, using parallel scan computation, a shuffle mechanism to mix batches, and a dynamic controller that adjusts batch sizes to trade off speed against quality. Across four benchmarks (AppWorld, Terminal-Bench, Formula, and FiNER), it ran up to 17 times faster than prior methods while matching or beating their accuracy, at comparable cost.
The real constraint on this kind of learning was never the idea. It was throughput. As more companies run swarms of AI agents on the same tasks, the ability to fold all of those parallel runs into one improving prompt, rather than processing them one at a time, is what turns agent trial and error into something resembling compounding improvement.
It is a scaffolding upgrade, not a smarter model. Combee does not teach an agent anything new in principle. It just lets the teaching happen faster and at higher volume, which for infrastructure-heavy AI labs may matter more than another benchmark point.