Context Is a Platform Capability Now

The article discusses the growing need for context as a platform capability in enterprise AI work. When developers begin working on real enterprise tasks, they go through a ritual of gathering essential information before making a useful prompt. This includes referencing deployment standards, runbooks, and explaining criticality tiers.

However, the agent used by the developer does not reliably accumulate this organizational truth on its own. Instead, developers must manually gather and provide the context themselves, which is time-consuming and prone to errors.

The author argues that internal platforms should serve both developers and their AI agents. Agents can remember more than they used to, but they still require organizational knowledge to function effectively. As more engineers adopt AI, the tax of manually gathering context grows, leading to inconsistent and outdated information being used. Platform owners face the challenge of making an organization's knowledge usable by its AI agents, which currently falls to developers assembling context independently.

To address this issue, the author proposes six properties for a trusted context layer: canonical (one correct answer per question), versioned (standards with history), fresh (measure staleness like uptime), attributable (source and owner association), accessible (machine-readable and queryable), and safe (inherit access controls of the described systems). These properties would help ensure that the context provided to AI agents is reliable, up-to-date, and easily accessible.