Concord keeps AI agents updated with changing workspace data
When Agent Context Goes Stale: Incoherence in Volatile Agent Context
Artificial IntelligenceOperating Systems
Summary
AI agents often rely on information they get from tools to make decisions, but this information can become outdated if the original data changes later. The authors identify a problem where AI models keep using old, stale data and make wrong conclusions. They propose Concord, a system that tracks where each piece of information came from, notices if the original data changes, and updates or flags the old information accordingly. This helps AI agents stay consistent with the current state of their environment and avoid outdated mistakes.
What this means in practice
- •For software development teams: Maintain synchronization between AI assistants and up-to-date project files to avoid errors caused by outdated file observations.
- •For data engineering teams: Ensure data-driven AI workflows detect and handle changes in underlying external data sources to prevent stale context usage.
Authors
Yingying Liu, Junzhou Fang, Chenxiong Qian
Abstract
Modern agents increasingly ground their reasoning in observations returned by tools, such as file contents read from a workspace. However, the data sources underlying these observations may later be modified by users, other agents, or external tools, while the model retains only the stale content in its context window. Existing agent runtimes provide little support for notifying the model that a previously observed fact has become stale, causing agents to reuse outdated observations and make incorrect claims about the current workspace state. We propose Concord, a context coherence framework that maintains the consistency between tool observation in agent context and the mutable sources from which they were derived. Concord links each observation to its source, detects source changes, and uses configurable handling policies to update, annotate, or suppress stale context before reuse. Concord is applicable across different agent runtimes and external resources, and can be easily extended to new runtime-resource settings. We implement Concord as a general framework, and instantiate a concrete use case to assess its effectiveness. We construct ConcordBench, where previously observed file contents become stale after subsequent edits. Across three evaluated frontier models, Concord produces answers consistent with the restored workspace state in all evaluated cases under these constructed conditions, matching the oracle on recover count for this benchmark, while using 46.4% fewer tokens than the strongest non-oracle baseline.