Databricks teaches Genie to handle multistep tasks and tell the CEO hard truths

The bigger the company, the less truth reaches the person in charge. A salesperson won't admit to the CEO that they botched a meeting. An HR manager won't say that the CEO's idea for improving employee satisfaction probably won't work. Databricks wants an agent to fill that gap: this week, the company unveiled an updated version of Genie.

What's changed in Genie

The main update is that Genie can now take on time-consuming, multistep tasks. According to the company, an executive can set a broad task and the agent will carry it out without supervision. Agents in this category were previously demonstrated mostly through software development examples; Databricks is now targeting the business side.

The agent gets its context from an ontology Databricks built on top of customers' information on its platform. That detail matters. An agent that doesn't understand what a company's data means quickly starts making things up. Here, it works with data already stored in Databricks and relies on that data's structure.

Whether this approach suits you depends on one thing: how much of your data already lives on the platform. If it doesn't, the ontology is considerably less valuable to you.

A demo featuring the CEO's real chats

Instead of a conventional product demo, Ali Ghodsi opened his own work agent and scrolled through recent conversations and charts. Here's what they showed:

• Genie analyzed the sections of an employee survey where the company received low scores. It then helped assess—and reject—remedies proposed by some senior executives.
• The agent summarized a major customer presentation that went badly, including features the customer refused to buy without.
• It prepared Ghodsi for a podcast interview by reviewing past appearances and suggesting likely questions.
• During a conversation with a journalist, the agent even suggested what Ghodsi should say. According to the TechCrunch writer, the advice was pretty good.

Ghodsi explained why he needs this by invoking Dunbar's number: a theoretical limit on stable personal relationships of around 150 people. At a small startup, the CEO knows everyone, and people feel comfortable delivering unpleasant news directly. At a company with 12,000 people, that no longer works. For him, the agent is therefore a source of unfiltered truth about what's going wrong. Ghodsi believes it can benefit any company with more than 150 employees—and there are plenty of those.

Costs, security and what to do about them

Ghodsi also uses Genie in his personal life. Meanwhile, he described alternatives such as OpenClaw, Instinct and Muse as questionable from a security standpoint. He believes Genie will do for businesses what AI, agents and harnesses—the scaffolding around a model that lets it use tools—are already doing for programmers.

The biggest fear when deploying agents is the token bill. Databricks' answer is a product called Unity Gateway. It lets users set usage limits and swap expensive models for cheaper open-weight alternatives.

What to keep in mind

Every example came from the head of the company selling the product. There are no independent measurements: no data on how often the agent makes mistakes on lengthy tasks or how much those tasks cost. The claim that it works without supervision remains the company's assertion. Neither pricing nor availability has been disclosed.

There are three practical takeaways for product managers and founders. First, an agent's value to an executive may lie in candid feedback that people won't deliver in person, rather than automation. Second, build in limits and the ability to switch to cheaper models when launching long-running autonomous tasks, instead of adding them after the bill arrives. Third, without a layer such as an ontology that explains the meaning of your data to the agent, autonomy won't get you very far.