LLM Agents Lose Track of Database Entities Across Multi-Step Tasks, Study Warns

Multi-step AI agents that interact with external systems face a subtle but significant reliability challenge: over the course of a workflow, they frequently lose track of which database entity they're supposed to be working with. Researchers from a new arXiv study have quantified this problem and tested potential solutions, revealing that the most intuitive fixes can actually make things worse.

The Problem: Entity Binding Drift#

When tool-augmented agents resolve a database entity at the start of a task—say, selecting the correct customer record—they should act on that same entity consistently throughout the workflow. But the research identifies two distinct failure modes:

• Binding drift: the agent selects the correct entity initially, then silently switches to a different one midway through
• Error propagation: the agent picks the wrong entity from the start and carries that mistake forward

Prior work had shown that agents misalign tools to entities roughly 24-26% of the time in single-step scenarios. This study traces what happens when those bindings persist across multiple steps in a controlled testbed: 200 workflows spanning four enterprise domains, tested against eight model backends ranging from smaller to frontier-scale LLMs.

The Intuitive Fix Makes Things Worse#

The most obvious response to entity drift is an "entity lock"—essentially telling the agent to persist the first binding it made and never reconsider. The researchers tested this approach and found it catastrophic. When they injected errors into the system to measure the impact, the entity lock amplified wrong actions from 907 to 2,746 (a 3.0× increase). On the worst-affected model (Claude Opus 4.5), amplification reached 8.5×.

Why? Because the lock does exactly what it's told: it faithfully carries an incorrect binding through every subsequent step, compounding the original mistake.

A Practical Re-Verification Solution#

The researchers evaluated an alternative: a lightweight LLM-based re-verifier that makes a single cheap additional model call to re-read the original instruction and double-check the entity binding. This simple approach reduced wrong actions by 79% (bringing the error rate to 0.21x), nearly closing the gap to oracle-level performance (0.20x).