Databricks updates Genie agent to run long, unsupervised jobs on customer data

Databricks has released an updated version of Genie, its AI agent, that can now carry out multi-step jobs requiring extended work rather than answering single-shot questions. According to TechCrunch, the company unveiled the upgrade on Wednesday and says business leaders can hand the agent a wide range of tasks that it will complete without supervision.

What changed under the hood

The upgrade rests on an ontology Databricks built to give Genie the context it needs from the data customers already store on the platform. That grounding is what lets the agent chain steps together across an organisation's own information — surveys, sales records, documents — for jobs that take longer than a typical chat exchange. TechCrunch reports that Databricks now pitches Genie as a persistent agent for business leaders rather than a simple query interface.

Ghodsi's unfiltered test drive

Rather than a scripted product demo, CEO Ali Ghodsi opened his actual agent for TechCrunch and walked the reporter through recent chats and charts. Among the examples he showed:

• It analysed the parts of an employee satisfaction survey where Databricks scored poorly, then helped him grade — and discard — remedies executives had proposed for those weak spots.
• It summarised a major sales pitch to a prospect that had gone badly, including the features the prospect said were required before buying.
• It prepared him for a podcast interview by reviewing his earlier appearances and drafting likely questions. Midway through the conversation, it even advised him on what he should say to the reporter, advice the writer found reasonable.

TechCrunch's central observation was that Ghodsi was using Genie to reach information he is unlikely to hear as the head of a 12,000-person company.

The Dunbar's number pitch

When this was pointed out, Ghodsi reached for Dunbar's number: the roughly 150-person ceiling British anthropologist Robin Dunbar proposed for how many stable relationships one person can maintain. In a small startup, Ghodsi argued, the CEO knows everyone and negative information reaches them directly. At Databricks' size, a salesperson will not confess a disastrous call, and an HR manager will not admit a satisfaction initiative is likely to fail. His view, as reported by TechCrunch, is that any company with more than about 150 employees — a large share of businesses — could get value from agents like Genie.

Ghodsi's own daily use of the product, classic dogfooding, is part of the sales story. He also told TechCrunch he runs Genie in his personal life rather than rival assistants he considers questionable on security, with the report naming OpenClaw, Instinct and Muse as the alternatives in question. His larger claim is that Genie can do for business work what agents and harnesses — the software that wraps around a model and lets it use tools — are already doing for programmers.

Handling the cost question

On the risk that persistent agents inflate corporate AI bills, Ghodsi pointed to Unity Gateway. According to TechCrunch, the product lets companies set usage limits and swap in less expensive open-weight models.

Why it matters

The update moves agents past the question-and-answer stage into unsupervised execution of long-running work — the threshold many enterprises have said they need before delegating real processes. Databricks' differentiator is the ontology: grounding the agent in data already sitting on its platform, territory generic assistants cannot easily reach. The Dunbar framing is marketing, but it names a concrete early use case in leaders using agents to bypass filtered internal reporting and surface what is actually going wrong. And the cost answer may matter as much as the capability: whether persistent agents get adopted will depend on governance tools such as usage caps and cheaper model options, not intelligence alone.