Researchers released a paper on Oct 6, 2026 describing SHIFT, a system that automatically creates the right combination of AI agents for each question. Instead of trying many setups at run time, SHIFT trains a small language model to predict which design will work best, then uses Monte Carlo tree search to assemble the agents. The approach was tested on 9,193 tasks across six benchmarks, ranging from math problems to document handling and general‑assistant work.
How does SHIFT perform?
With a Gemini 3.5 Flash model as the execution engine, SHIFT achieved an average accuracy of about 80%, beating 17 competing methods that include plain prompting, prompt‑optimization tricks, and other workflow searches. The strongest competitor fell short by 7.2 percentage points. A cheaper configuration of SHIFT still outperformed every baseline while using 32% fewer tokens, meaning it required less compute.
Why the joint design matters
The authors show that picking the communication structure, instructions, and tools together works better than picking any one of those pieces alone. The joint choice improved results by up to 9.1 points compared with using only instructions or only tools. Their learned value function also helped pick more accurate, cheaper harnesses from a pool of candidates.
Why it matters
For developers building AI assistants, SHIFT promises higher performance without needing to hand‑craft agent setups for each new request. It also reduces compute costs, which can lower cloud expenses and make richer multi‑agent services more affordable for end users.
