Adopt Databricks’ three Cs thesis, but don’t bet the stack on one vendor

This HFS Highlight is for CIOs, chief AI officers, and data and AI leaders evaluating how to act on Databricks’ context, cost, and control thesis without overcommitting to a single vendor.

Fix the three Cs of context, cost, and control, and you unlock enterprise AI. That’s the thesis emerging from Databricks’ Data + AI Summit in San Francisco, held on June 15–18. CEO Ali Ghodsi told the room what many enterprise leaders have already concluded: “AI does not have an intelligence problem, it has a context problem.”

The diagnosis is right, and data and AI leaders should act on it. But they must do so with caution. The same vendor that champions “no lock-in” is assembling one of the most complete agentic stacks in the market. Few enterprises should be willing to hand their entire data and AI estate to a single supplier. Our advice: pilot, but carefully govern the dependencies you may create.

Move spend and focus from models to context, cost, and control

Databricks frames enterprise AI around context, cost, and control. This mirrors what HFS sees as stalling enterprise programs: the fragility of agentic workflows when operating without an understanding of the enterprise’s internal language and ways of working, runaway token costs, and governance issues.

Recognizing these barriers should reset your priorities. Stop chasing the latest model, and invest instead in the context and governance layers that actually determine production quality and cost.

On context, Genie Ontology and its OntoRank index (see Exhibit 1) pre-compute a knowledge graph across enterprise data assets, effectively auto-generating your context to spare agents the costly “random walks” of discovery at query time. The logic is sound and worth piloting, but it challenges the quality claims. This context is built on data and is unlikely to surface the tribal knowledge and tacit understanding of how work gets done. Treat this as unproven until you test it against your own data.

On control and cost, Unity AI Gateway (see Exhibit 1) is a one-plane capture of budgets, observability, identity, and guardrails across every agent, model, and MCP server. It answers concerns about runaway spend and oversight gaps. This must become table stakes for any agent program, whether procured from Databricks or not. As an open source, it could stand as a reference architecture for those assembling their own.

Exhibit 1: The Databricks’ Data + AI 2026 Summit revealed the scale of its enterprise AI ambitions

Source: HFS Research, 2026; based on Databricks Data + AI Summit keynotes and sessions

Treat Omnigent, the new “meta-harness,” as your cue to demand interoperability before scaling agents

“Harness” is going to be one of the words of the year. Databricks announced Omnigent, an open-sourced meta-harness it hopes the world will adopt and a signal to AI and data leaders that interoperability is essential before multi-agent, cross-ecosystem agentic systems are deployed.

A harness is the software layer that sits around an LLM and turns it into something useful. It connects a model to files, tools, permissions, interfaces, workflows, and the outside world. Claude Code has one. Codex has one. Every custom agent effectively has one. We are accumulating harnesses almost as quickly as we’re accumulating models, and that creates three growing enterprise headaches:

• Build: Teams increasingly deploy multiple agents together, using different models, agent frameworks, and strengths. Most harnesses operate as islands, making it difficult to orchestrate agents across tools and vendors.
• Handoffs: Many engineers (and a rising number of knowledge workers) have reached the stage where our desktop is a collection of agent windows, Slack, and Team threads and documents, with a surprising amount of copy-and-paste acting as the integration layer. Context is trapped inside individual agent sessions, making handoffs between people and agents hard.
• Consistency: Security, cost management, and governance become increasingly difficult when every agent comes with its own way of handling permissions, policies, and oversight. Prompt injection, unintended actions, and runaway spend are all management concerns, not just technical ones.

Omnigent wraps different agents and presents them through a shared interface to ease composition, real-time collaboration, and contextual, cost-aware control. The intent is to support orchestration, governance, and collaboration across a sprawl of otherwise incompatible agents.

As a leader, use Omnigent as an opportunity to insist on a layer that lets you compose, switch, and govern underneath. Treat it as a candidate standard, but judge it on adoption beyond Databricks, especially as it does not require Databricks tooling.

Omnigent’s value depends on it becoming a widely used open ecosystem. Databricks’ decision to open source it rather than monetize it is a good sign. But you would be wise not to architect around it until the ecosystem materializes.

The Bottom Line: Pay down your context and governance debt, but don’t put all your eggs in one basket to achieve this.

Act on the Databricks thesis and start paying down your context and governance debts today. Make agent cost control mandatory and built in as part of your AI tooling.

Where Databricks offers open, interoperable pieces (Unity AI, Gateway, Omnigent), be prepared to adopt aggressively. But on proprietary offerings such as Genie Ontology, Customer Lake, and Lakewatch, pilot with an eye on quality and cost gains.

Databricks may come to you with an all-in-one ambition, but the sheer breadth of this stack should heighten scrutiny rather than earn blanket commitment.