New Trick Trains AI Agents 5x Faster, Slashing Costs

Agentic AI is the kind of AI that doesn't just answer questions — it can carry out tasks, like booking a flight or managing an inbox. But training these AI agents is incredibly expensive and slow. The training process involves showing the AI millions of examples and adjusting it through trial and error, which takes enormous computing power.

Researchers noticed something smart: in large-scale training, many of those examples start in exactly the same way. For instance, every time the AI begins a task, it first reads the same instructions and processes the same context. The new system, called psRL, shares that work across many training samples at once. Instead of each example doing the same initial steps separately, they reuse the result — like multiple chefs prepping vegetables from one shared cutting board.

This sharing sounds simple, but it requires clever coordination across many computers. psRL manages the memory and workload across graphics processors (GPUs) so that no machine sits idle or gets overloaded. In tests using real production data, psRL made training up to 5.2 times faster than existing systems. That's not just a small improvement — it could mean the difference between spending months and spending weeks on a single training run.

The practical payoff? Faster, cheaper training means AI companies can iterate more quickly, experiment more safely, and potentially pass savings on to consumers. While this paper is about the underlying plumbing of AI, it's a vital piece of making AI agents more common in daily life.