Sponsored by: Atlan | The Enterprise Context Layer: Why Agents Fail — and the Fix — Demystified and Demoed

As organizations move toward AI agents and agentic systems, differentiation is shifting from raw model capability to how well agents understand a specific business. The Databricks Platform gives teams the unified foundation to capture and operationalize that context at scale.

The session argues that performance equals intelligence plus context. Intelligence has scaled, yet most companies still pilot one-off demos because the tacit context that lives in dashboards, Slack threads, SQL, and people's heads remains uncaptured. Turning that into machine-usable form is what separates a prototype from production.

From raw intelligence to contextual performance

High-intelligence systems that lack business grounding operate in a low-context danger zone. Effective agents must know not only what a metric is and where it lives, but also when a pattern is seasonal, which entities and store types matter, and what a user is allowed to see. That leap from answering what to explaining why is the frontier for enterprise AI.

Context becomes a moat as intelligence commoditizes. How a company defines its terms, does work, and enforces norms is unique, and the value compounds as agents learn and apply that context consistently across use cases.

What an enterprise context layer includes

The conversation frames a context layer as the organization's living brain that converts knowledge, expertise, and norms into machine-usable artifacts. Three reproducible units anchor it:

• AI-ready data and a knowledge graph, so agents know what data means, how it connects, and how it is used in practice.
• Semantics and ontology, so metrics are defined, related, and disambiguated across domains, and entities map to each other coherently.
• Skills, an open, reusable format for encoding procedures, expertise, and norms, so agents can follow how work actually gets done.  
   

Enterprises face recurring challenges building this layer: bootstrapping context for open-ended questions, governing context that changes as the business changes, and avoiding agent sprawl where each platform learns its own truth. Treating context like a shared, versioned repository prevents fragmentation and makes it portable across AI/BI Dashboards.

Mining, governing, and activating context

Rather than documenting everything from scratch, much of the business brain can be reverse-constructed. Lineage from systems of record through the warehouse to BI, combined with SQL history and existing semantic logic, can generate accurate descriptions, filters, common business questions, metrics, and ultimately an ontology that links logical entities to physical tables. Quality compounds when these signals are combined.

A context development lifecycle pairs AI with targeted human judgment. AI proposes context, builds tests, and reviews conflicts. Humans resolve ambiguity and approve decisions that require intent, such as choosing between two valid metric definitions used by different teams. Before rollout, simulations generate likely business questions and evaluate answers, and the system suggests precise changes to improve accuracy. One-click delivery pushes approved context into targets like Genie, Claude, OpenAI via MCP, or even deep agents in LangChain, keeping the repository open, Iceberg-native on S3, and portable across tools.

Centralizing context mitigates agent sprawl. A single open repository feeds every agent and platform, so learning in one place propagates, rather than creating competing versions of truth.

Human on the loop and compounding learning

The aim is not to keep humans in every step, but to keep them on the loop for the few decisions where intent matters. Every interaction produces traces that reveal patterns and gaps. Agents can mine those traces to suggest context updates, while humans steer only where needed. Customers reportedly launch agents at roughly 70 percent autonomy, then climb toward 90 to 95 percent as feedback loops mature.

Make context your portable moat

Enterprises will not scale agents by hard coding one-off prompts and schemas. They will scale by operationalizing context as a first-class artifact: mined from existing systems, governed like code, improved by traces, and deployed everywhere. Keep it open, keep it portable, and keep it yours.