AI agents silently drift on entity bindings 18% of the time

A new arXiv paper by Rahul Suresh Babu and Shashank Indukuri investigates a critical flaw in tool-augmented language-model agents: binding drift. These agents often resolve an entity once and then reuse that binding across multi-step workflows. The researchers formalized two separate failure modes — binding drift (correct at step 1, wrong later) and error propagation (wrong from the start, carried forward) — and tested them on a controlled testbed of 200 workflows spanning 580 entity-binding-scored steps across four enterprise domains and eight model backends, from small to frontier models.

Key findings include that a naive entity-lock fix (persist the first binding) actually made things worse: under controlled error injection, it amplified wrong actions from 907 to 2,746 (3.0x), and hit 8.5x on Claude Opus 4.5. In contrast, a practical LLM-based re-verifier — a single cheap second model call re-reading the original instruction — reduced wrong actions by 79% (0.21x), closing the gap to within 1 percentage point of an oracle upper-bound. In natural settings, baseline agents already drift on 18% of eligible workflows, with per-step error rates rising across steps. The paper shows that persistence and re-verification are not interchangeable and that a simple re-verification strategy nearly matches oracle recovery.