New Trick Makes AI Agents Up to 32% Faster by Guessing Ahead

When you ask an AI agent to handle a multi-step job — research a topic, fill in a form, check a spreadsheet — the software pauses at every fork in the road, waiting for the model to pick a direction. Nothing downstream can start until that choice is made. Researchers call this the "branch-resolution barrier." It's a bit like a delivery driver sitting at an intersection with the engine off, refusing to move until the route is chosen.

A team of four researchers — Junyi Shen, Noppanat Wadlom, Zhengyuan Su and Yao Lu — built DynBranch to fix that. The trick is giving each undecided fork a stable name, so two things can happen at once. The system quietly starts likely next steps while the model is still deciding, and it recognizes when a finished chunk of work matches something done before and simply reuses it. A controller only allows this extra work when the savings outweigh the added strain.

The results: on four realistic agent workloads running Qwen3-32B, an open model, across four high-end Nvidia H200 chips, average waiting time dropped by up to 32 percent compared with each workload's best previous system, and by 46 to 66 percent against a version that reused nothing. The speedup held on much cheaper hardware too — an 8-billion-parameter model on a single consumer RTX 4090 card. Crucially, the final answers stayed exactly the same; only the speed changed.

The catch: this is a research paper, not a product you can download today. Speculation also costs money when the guess is wrong, which is why the controller is cautious. And the gains are biggest for repetitive agent workflows that hit similar steps again and again — a one-off, unusual task may see little benefit. Still, as companies run agents at scale, shaving a third off waiting time adds up fast.