Fine-grained error attribution on frontier agent trajectories reveals two systematic bottlenecks:
• Deep Research Failures: Dominated by Retrieval Drift (32%–41%)—agents reach the relevant research neighborhood but latch onto cited baseline papers or adjacent works instead of resolving all constraints.
• Wide Research Polarization: Claude-Opus-4.6 exhibits 85.3% Unconstrained Expansion (over-collecting unverified papers at the cost of precision), whereas Seed-2.0-Pro suffers 33.3% Premature Termination (quitting after retrieving only a few candidates).
Agents systematically lack stateful evidence tracking and the capability to delineate precise scientific concept boundaries.