AI agents that read a file once and never check again are getting a fact-checker.
Researchers built a system called Concord that tracks the files, and other mutable data sources, an AI agent has already looked at. When those sources change after the agent has read them, Concord flags the agent's copy as stale and either updates it, annotates it, or blocks it from being reused, depending on a configured policy. The team tested Concord on a new benchmark, ConcordBench, where file contents get edited after an agent has already read them. Across three frontier models, Concord caught every case of stale data in the tests, matching a theoretical best-case baseline, while using 46.4 percent fewer tokens than the next-best approach.
Most agent frameworks treat whatever a tool returned as permanent truth, even after the underlying file has moved on. That gap is easy to miss in a demo and expensive in production, since an agent that confidently reports on a file that no longer exists is worse than one that admits it doesn't know.
It's a narrow fix for a problem that will only grow as agents juggle more tools, more collaborators, and more moving parts. That makes it plumbing, not a breakthrough, but plumbing that's currently missing.