Sankar defines enterprise contextContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task. as three connected resources: factual knowledge about the business, expert playbooks for using that knowledge, and organizational norms such as permissions and approval boundariesGovernance defines responsibilities, decision rights and controls for how a system or organization is operated.. A capable model without this context cannot reliably understand company-specific language, metrics, workflows, or constraints.
Atlan first built narrow agents for individual jobs, then shifted toward general-purpose agents that draw from shared context. Marketing teams eventually accumulated hundreds of skillsAn AI agent skill is a reusable package of instructions, resources, and tool guidance for performing a bounded kind of work. and dozens of agents, exposing familiar software problems such as dependency drift, unclear ownership, secret management, quality control, and duplicated instructions.
The proposed context layer works like a GitHub for operational knowledge. Teams can version skills and facts, assign maintainers and approvers, inspect dependencies, test quality, and reconstruct answers through SQL, vector retrievalRetrieval is the process of selecting relevant stored information and returning it to an AI system for the current task., MCP tools, or hybrid assembly. This makes context reusable across different models and agent products instead of trapping it inside one application.
Sankar also describes a learning loopAn AI feedback loop occurs when AI outputs or their consequences become inputs that influence later model or agent behavior. that traces what an agent retrieved, what it attempted, and how a human corrected the result. Those traces can improve the underlying context over time. The resulting repository becomes valuable company intellectual property because it captures how the organization actually reasons, decides, and executes.
Watch on YouTube
