In June 2025 Andrej Karpathy wrote that he preferred the term context engineering to prompt engineering, and described it as the craft of filling a model's context window with the right information for the next step.
The remark traveled, largely because it named something most teams building with language models had already felt: the model had stopped being the weak part. A capable model given the wrong page of the company wiki produces a fluent, well-organized answer about a plan the company dropped in March.
An AI context layer is the answer companies are converging on. It is the system that holds current, structured knowledge about the business (its goals, the projects meant to deliver them, who owns each one, what changed this week) and hands the relevant part to any model or agent that asks. One layer serves every agent, so the sales assistant and the planning agent start from the same account of the company.
What does an AI context layer hold?
Arguably the term is easiest to grasp through what it replaces. Today most companies give their AI context in two ways. Someone writes an instruction file, a page of background that the agent loads at the start of each session, and someone connects a search tool so the agent can pull documents on demand. Both help.
However, an instruction file describes the company on the day its author wrote it, and a search tool returns whichever documents happen to match the question, including the superseded plan and the draft nobody approved. We go through those failures one by one in Why AI Agents Get Your Company Context Wrong.
A context layer differs on two counts. It is structured: a goal is an object with an owner and a due date, linked to the projects that serve it, where a document only mentions those things in prose. It is also current. When an engineer moves a Jira ticket, or a director says in a meeting that a launch is waiting on legal, the layer changes the same day, and nobody had to write that down for the AI's benefit.
Why does strategy need its own context layer?
Most writing about context layers concerns data: metric definitions, table lineage, which revenue number is the official one. That work matters a great deal to an analyst's agent.
An agent helping a team lead decide what to do on Tuesday needs something else, namely the plan and the state of the plan. Which goals did leadership commit to this year? Which projects serve them, and which of those have gone quiet? Does the task in front of this person connect to any of it?
No single tool holds the answer. The strategy sits in a slide deck, the projects in Jira or Asana, the decisions in Slack threads and meeting transcripts, the numbers in a spreadsheet. We spent seven years running strategy execution programs in more than 300 organizations, and the pattern held in nearly all of them: leadership assembled its view of execution by hand, from status updates that people wrote late and rounded upward (the mechanics are in The Green Dashboard Problem).
An AI agent pointed at the same sources inherits the same blind spot, which is why AI pointed at status reports produces a confident summary of a summary.
How does Vindaris build the context layer?
Vindaris is the AI context layer for strategy execution. Its Work Graph maps the strategy to the goals, KPIs, projects and tasks under it, and it analyzes work activity across the tools your teams use, such as Jira, Slack, Microsoft 365 and meeting transcripts. From that activity it derives an observed state for every item and flags where effort diverges from the plan, weeks before the next review would show it. Your teams write no status updates for it.
The same graph serves people and machines. An employee sees their highest priority for the day. An agent connected over the Vindaris MCP server can ask which projects under a goal are blocked and get an answer with the evidence attached, down to the thread in which someone said so.
The structure is public, too. We published the Work Graph schema with its object types, link types, signal kinds and alert definitions, so you can inspect what an agent will work from before you connect one.
Does the choice of model still matter?
None of this makes the choice of model irrelevant, of course. A stronger model draws better conclusions from the same context. However, the gains now arrive unevenly: every lab ships a better reasoner every few months, and no lab ships knowledge of your company. That part you supply, once, in a form all your agents share, or you supply it again in every prompt.
If every agent in your company can soon reason equally well, what will separate the useful one from the merely confident one, other than what it knows about your plan this morning?
FAQ
Is an AI context layer the same as a knowledge base? No. A knowledge base stores documents for people or models to search. An AI context layer holds structured objects such as goals and projects with their owners, links and current state, and it updates as work activity arrives from connected tools.
How is a context layer different from a system prompt or an AGENTS.md file? An instruction file is written once by a person and loaded into every session. A context layer is maintained by a system and queried on demand, so an agent gets the state of the company as of today.
Who needs an AI context layer? Any company that runs more than a few AI agents or assistants on internal questions. Without a shared layer, each agent assembles its own picture of the business and the answers disagree.
Does Vindaris work with our own AI agents? Yes. Vindaris runs an MCP server, so Claude, ChatGPT, Claude Code and other MCP clients can query the Work Graph within the permissions an admin sets.