Databricks 升级 Genie 智能体 可无人监督完成多步任务

Databricks on Wednesday unveiled an updated version of its AI agent Genie, which can now perform multi-step jobs that take extended work to finish. It does this through an ontology that Databricks built to give Genie the context it needs from the information that customers store on the Databricks platform.

Corporate leaders can now assign the agent a wide range of tasks, and it will complete them unsupervised, the company says. To prove the point, Databricks CEO Ali Ghodsi showed me how he uses it internally. Instead of a product demo, he fired up his actual agent and walked me through his most recent chats and charts.

He showed me how it analyzed the parts of an employee satisfaction survey where the company didn’t score well, and how it helped him rate (and discard) fixes for those low scores that some executives suggested. He showed me how it summarized a big sales pitch to a prospect that went poorly — including the features the prospect said it needed before buying.

It also prepped him for a podcast interview by reviewing his previous appearances and developing likely questions. While we talked, he even had it tell him what he should say to me. (And its advice on that wasn’t bad.)

The fascinating part was that Ghodsi was using Genie to find the stuff that he’s now unlikely to hear as the big boss. It was a source of unvarnished truth about what wasn’t going well that he needed to address.

“Have you heard of Dunbar’s number? The magic number is 150,” he said when I pointed this out. Named for British anthropologist Robin Dunbar, it’s theoretically the maximum number of people with whom humans can maintain stable relationships. In a small startup, a CEO knows everyone, has direct relationships with them, and they feel safe to tell him negative info straight up.

With 12,000 employees, Ghodsi is the first to admit he can’t know everyone. At that size, a salesperson isn’t likely to confess a disastrous sales call to the CEO. An HR manager isn’t likely to confess that their idea for increasing employee satisfaction will probably fail.

 var playerInstance_jwplayer_6af482fde1097 = jwplayer( "jwplayer_6af482fde1097" ); playerInstance_jwplayer_6af482fde1097.setup({ playlist: "https://cdn.jwplayer.com/v2/media/YS4q5560", });   

“Anyone who has more than 150 employees, which is a lot of companies, a lot of people can get value out of this,” he said of AI agents generally and his product, Genie, specifically.

Ghodsi’s own use of Genie (often called dogfooding, from the phrase “eating your own dogfood”) is one reason why Databricks is now pitching it as an always-on agent for business leaders. He says he also uses it in his personal life, rather than a security-questionable alternative like OpenClaw or Instinct or Muse.

He believes Genie can do for the business side what AI, agents and harnesses (the software that wraps around an AI model and lets it use tools) are doing for programmers. As for the risk that agents will ratchet up corporate AI costs, Ghodsi has an answer for that, too. He said costs can be controlled through Databricks’ Unity Gateway product. It allows companies to set usage limits and to swap in less expensive open-weight models.