After a first day defined by AI agents that fix data pipelines and make decisions inside enterprise systems, Data+AI Summit 2026 reached Wednesday, June 17, in San Francisco with a complementary and perhaps even more urgent message: having an AI agent is no longer the challenge. The challenge is having dozens of them, spread across engineering, marketing and security, and still keeping everything under control. Databricks devoted much of day two's announcements to this issue, introducing ways to combine, secure, govern and even visually customize the work of these agents.
One layer to command every agent at once
Until now, every AI agent came with its own way of working, its own way of receiving instructions and its own history, isolated from the rest. In practice, a professional might have five or six different agents open at once, copying and pasting information between them, with no layer that sees the whole picture. Databricks introduced Omnigent, open-source software that sits as a layer above the agents a company already uses, whether in-house coding assistants or those from other vendors. This layer lets teams combine different agents into a single workflow, apply security and cost rules to all of them at once — such as pausing an agent after it spends a set amount — and share work sessions in real time with teammates, like a collaborative document. The logic mirrors the evolution we've already seen in server management: companies stopped looking after each machine individually and began managing entire fleets from a central console. According to Databricks, AI agents are ready for the same shift.
Marketing also needs a real-time brain
CustomerLake, one of the most talked-about announcements on day one, got a deeper explanation the next day of why it exists. Databricks argues that consumer behavior has already changed: today, AI agents routinely search for flights, compare prices and even complete purchases on a person's behalf, all in a matter of seconds. That makes traditional marketing tools — built for batch campaigns decided by humans and updated every few days — obsolete. The company's answer lies in the concept of the Agentic CDP, of which CustomerLake is the first practical application: a customer data platform designed for the agent era, embedded directly in the company's own data foundation instead of running as a separate system that must be kept in sync. In practice, if a customer's flight is delayed and they're traveling with their kids, the system can react immediately — offering lounge access, for example — instead of sending a generic loyalty email days later. It's the difference between a concierge who already knows what's happening with a guest and a marketing department that only learns about it in the next campaign.
Genie Code gets a command center and now understands AI models too
Genie Code, the Databricks agent already used by 90% of the company's customers, gained a full-screen command center designed for people juggling multiple data tasks at once. Instead of tracking each task in a separate tab, users see in a single panel the progress of every workstream, what's already done and what's still awaiting a human decision — like swapping dozens of open browser tabs for a control tower that shows everything at once. Databricks also announced that Genie Code will soon be able to work autonomously on a schedule, running tasks overnight and delivering results ready for review the next morning — like an extra shift the team doesn't have to supervise.
The same tool now also has a deep understanding of the AI model-building process, from data preparation to monitoring once the model is in production, because it now has access to the experiment history and governance rules each company keeps in its own data foundation. This keeps the agent from behaving like a generic assistant that merely guesses what the company needs, and turns it into someone who already knows the company's history. Databricks also expanded the infrastructure behind this effort: it now offers on-demand GPU access for model training, without companies having to build their own compute infrastructure — the equivalent of buying electricity instead of building your own power plant — and strengthened its real-time serving layer, now able to handle more than 300,000 requests per second with just a few milliseconds of latency, enough to support tasks such as fraud detection and personalized recommendations at the exact moment a customer is making a decision.
Security and governance get an open ecosystem
Another set of announcements addressed an issue that has become unavoidable as AI agents start acting inside enterprise systems: who ensures they do only what they're supposed to? A day after unveiling Unity AI Gateway, its governance layer for AI models and agents, Databricks announced the first major ecosystem of partners connected to that layer, bringing together well-known security vendors such as Palo Alto Networks and CrowdStrike and identity management providers such as Okta, now adapted to monitor the behavior of agents, not just people. The logic is simple: if an AI agent gains access to internal systems as if it were an employee, it also needs a badge, clear rules about where it can go and someone watching what it does — exactly as with anyone the company hires.
Databricks also strengthened the platform's own security and compliance infrastructure, with automatic identity for more enterprise login providers, a way to expose only specific tools — such as a dashboard or a data chat — to external users without opening up the entire company environment, and new compliance certifications across cloud providers, including standards used by hospitals, government agencies and heavily regulated industries. For companies already facing strict audit requirements, this expansion removes an entire layer of manual compliance work.
Finally, Databricks returned to the story around LakeWatch, its security tool that tackles cyberattacks the same way it tackled enterprise data years ago: by unifying everything in a single open repository instead of scattering information across dozens of tools that don't talk to each other. The rationale is that attacks themselves are now automated by AI, and the average time between discovering a vulnerability and exploiting it has dropped from more than three weeks to less than two days.
Dashboards join the brand identity conversation
In a quieter but highly practical announcement, Databricks showed how to customize the look of AI/BI dashboards to match a company's visual identity, with its own fonts, colors and layout, instead of the default look every other company using the tool would have. More importantly, this customization can be applied once, at the workspace level, and automatically applies to every dashboard created afterward.
What this means for decision-makers
Taken together, day two's announcements make it clear that the enterprise AI conversation has entered a new phase. It's no longer enough to have one AI agent working well in a single process; the question leading companies are asking now is how to organize, secure and trust a growing number of these agents, spread across areas as different as data engineering, marketing and information security.
For sectors that deal daily with sensitive data, regulatory requirements and business-critical processes, the message is direct: data and AI maturity is no longer measured by the number of models in production, but by the ability to govern — with clear security and identity — everything those models and agents do inside the company.