Ask the same agent the same question about your company on Monday and again on Thursday, and there is a fair chance you get two different answers. That sounds like a complaint about the model. In most cases the model did exactly what it should with what it was given, and what it was given had changed between Monday and Thursday, or had not changed since April.
Why do instruction files such as AGENTS.md go stale?
The first source of context most teams set up is a file: AGENTS.md, CLAUDE.md, a system prompt, a pinned project brief. Someone capable writes down what the company does and what its priorities are, and for a few weeks the agent performs noticeably better.
However, the file records the company as its author understood it on one afternoon. The reorganization in September and the goal the board replaced in October never reach the file unless someone remembers to edit it, and editing a context file is exactly the sort of chore that loses to every deadline.
Agent memory looks like the remedy, and arguably it makes things worse. Memory features keep what came up in recent sessions, so the agent recalls last week's correction and forgets the decision from six months ago that still governs the project. It ends up holding a view of the company weighted toward whatever you discussed most recently.
Why does the same question get different answers?
The second source is retrieval. You connect the agent to SharePoint or Google Drive, or to several tools over MCP, and let it search when it needs to know something. Many teams have by now tried the blunt version of this and loaded every document they own into Claude or ChatGPT, on the reasoning that a model with access to everything will work out what matters.
It does not. A document store holds every version of the truth your company has ever written down, the strategy deck from the offsite next to the revision that replaced it two weeks later, and search ranks them by resemblance to the question. It has no way to know which one your leadership team still stands behind.
Retrieval is also unstable. Which tools the agent calls and which passages come back depend on small differences in phrasing, so two runs of one question assemble two different pictures. For a coding task that variance is tolerable, because the tests settle it.
For a question such as whether the Q4 launch is on track, nothing settles it, and you cannot build a weekly routine on an answer that changes with the wording. (For where retrieval does work well, see Context Layer vs. RAG vs. Enterprise Search.)
Why can't an agent find the current state of a project?
Suppose you solved both problems, with a diligent owner for the instruction file and a clean document store. The agent would still miss the most useful part, because the current state of a project rarely exists as a document. It exists as a ticket that has sat in review for nine days and as a thread in which an engineer mentioned waiting on a vendor.
The written summary, if one exists, is the status report, and status reports run late and optimistic for reasons we covered in The Green Dashboard Problem. People dislike writing them, so they write them quickly and round upward.
An agent that plans from the status report therefore plans from the most flattering account available. It will tell you, fluently, that the launch is on track.
What does a maintained context layer change?
What would it take for an agent to give the same correct answer on Monday and on Thursday? It would need one maintained account of the plan and of the activity around it, which every agent queries and which changes when the activity does. An AI context layer does that job.
Vindaris builds this for strategy execution. Its Work Graph analyzes work activity across the tools your teams use, such as Jira, Slack, Microsoft 365 and meeting transcripts, and compares what people spend their time on with what the strategy calls for.
It flags where the two diverge, weeks before the next review would show it. Agents reach the same graph over MCP, so the answer about the Q4 launch comes from this week's tickets and threads, with the sources attached, whoever asks and however they phrase it.
The instruction file will not disappear, and it should keep the things that seldom change, such as tone and house conventions. Everything that moves belongs somewhere that moves with it.
How much of what your agents told you last week rested on a page someone last edited in spring?
FAQ
Why does my AI agent give different answers to the same question? The agent searches your tools afresh on each run, and small differences in wording change which documents come back. Two runs therefore reason from two different sets of sources.
Is a bigger context window the fix? No. A larger window lets the agent load more documents, including outdated ones. It cannot tell the agent which version your leadership team still stands behind.
How often should we update AGENTS.md or CLAUDE.md? Keep those files for things that seldom change, such as conventions and tone. Anything that changes weekly, such as priorities and project state, belongs in a system that follows your work tools.
What fixes inconsistent company context for AI agents? One maintained source that every agent queries. Vindaris provides this for strategy execution: its Work Graph analyzes work activity across your tools and gives each agent the same current account of the plan.