Two AI agents taking turns on the same task beat one agent working alone - and did it in a quarter of the time.
Researchers studying long-horizon machine learning engineering agents found that a single agent loses steam the longer a session runs: reasoning quality declines as progress slows, a pattern they call long-horizon deceleration. They also found that the custom tool-and-memory harnesses built to prop up these agents lose their edge as the underlying models improve, a problem they call harness staleness. Their fix, Humanize MLE Agents, alternates two agents on a shared workspace, caps each session at a fixed number of experiments, and wipes the context clean between handoffs - no extra harness, no injected expertise required. On 75 Kaggle tasks from the MLE-bench benchmark, it hit a 78.2% any-medal rate, 8 points ahead of the strongest open-source baseline, using a quarter of that baseline's time budget.
The result suggests the bespoke scaffolding many teams build to keep long-running agents on track - memory systems, tool libraries, accumulated playbooks - may be solving the wrong problem. A simple reset-and-handoff between two agents produced consistent gains, averaging 6.4 points over the underlying baselines across five different agent pairings, and even carried over to an unrelated kernel-optimization task. That's a cheaper, more portable fix than most harness engineering, assuming it holds up outside benchmark conditions.
It's a strange answer to a problem AI labs mostly tried to solve by adding more engineering: sometimes the fix for an agent running out of steam is just to let a fresh one take over.