Databricks Bets the Hard Part of Agents Is the 99% Around the Loop

On June 16, at its Data + AI Summit in San Francisco, Databricks announced the expansion of Agent Bricks into a comprehensive agent platform for developers. The company said more than 100,000 agents have been built on it since launch, now processing over a quadrillion tokens a year, with customers including AstraZeneca, 7-Eleven, Fox Corporation, and Block. The framing is the most useful part: "the core agent loop is just 1% of the work. The other 99% is the hidden technical debt of agentic systems" — token capacity, deployment, security, evaluation, monitoring, context, and sharing. The platform organizes around three problems: Choice (multiple frontier and open models, from OpenAI, Anthropic, Gemini and Qwen to a newly added Kimi and a Grok partnership), Context (a "Genie Ontology" that learns business semantics, MCP support inside Unity Catalog, and managed agent memory), and Control (a new Unity AI Gateway to discover, govern, budget, and monitor every agent, model, and tool in one place).

For operators, this is a useful corrective to the demo culture around agents. The flashy part — a model calling tools in a loop — is the part you can stand up in an afternoon. The reason an agent gives a wrong answer in production is almost never the loop; it is that it retrieved stale context, couldn't tell which table was authoritative, or had no budget and no audit trail. Databricks is right that the data estate is "littered with missing or misleading information," and that the fix is semantic, not cosmetic. The honest read for a WMS/ERP shop: before you buy an agent, check whether your system of record can answer "which inventory table is authoritative" or "what counts as a late shipment here" — because that semantic layer is the actual moat, and it stays yours regardless of which model is fashionable this quarter. The risk in any single-vendor platform is the familiar one, with governance, memory, and ontology all converging on one stack. But the diagnosis is correct: spend on context and control, not on the loop.