Baseten co-founder and CEO Tuhin Srivastava joins Altimeter's Apoorv Agarwal for a wide-ranging conversation on AI inference: how Baseten went from building early machine-learning tooling to focusing on inference, why Tuhin frames training as a function of inference, and how he thinks about compute strategy, cost, and helping customers own more of their own stack.
The conversation covers the shift from renting to owning intelligence, when it makes sense to move off closed-model APIs, what it takes to serve models in production, how open and closed model usage splits for the most sophisticated...
What Matei Zaharia said
Written from the verified transcript and checked against it. Every figure links to the moment it was said.
Matei Zaharia, CTO of Databricks, introduced Omnigen, a meta harness that runs on top of existing agents like Cloud Code and Codex, providing common features such as collaboration, security, and the ability to swap or combine agents. He explained that the biggest challenges engineers faced were collaboration and security, which led to the need for a centralized layer. Zaharia highlighted features like setting budgets on sub-agents, sharing sessions for code review, and using markdown plans with comments. He discussed composability patterns, including orchestrator and sub-agent setups, and noted that open sourcing Omnigen benefits from community contributions, similar to Apache Spark and MLflow. He mentioned a benchmark showing Opus performs the same with the pie harness as with Cloud Code but costs about 2x less. Future plans include adding built-in agents, an auto research agent, and improving security and cost management policies.
Key takeaways
Omnigen is a meta harness that runs on top of existing agents, providing collaboration, security, and the ability to swap or combine them.
Zaharia stated that Opus performs the same with the pie harness as with Cloud Code but costs about 2x less, based on an internal benchmark.
Omnigen allows setting budgets on sub-agents, a feature no existing coding agent offers, preventing surprise bills.
Open sourcing Omnigen enables community contributions, with supported coding harnesses growing from 3 to 20 since launch.
Future plans include an auto research agent and making it easy to write custom agents in YAML.
8:35... of the underlying model.' Is there a performance boost you can get from having a meta harness in the same way you can get a performance boost from an individual harness on an agent or is there less of a performance story around Omnigen?
It's a layer that really benefits from many companies, many people contributing because there's a long tail. There's so many different types of agents.
Awesome. We're back here for the next episode of First Pass with Matei from Databricks to discuss Omnigen and how the evolution of harnesses and meta harnesses with respect to agents is evolving. Matei, thanks for being with us.
So, talk to us about Omnigen. Why did you build it? Well, first, what is it? Why did you build it? How did some of the internal work that you were doing lead you to building this? Just talk to us more about what Omnigen is.
Yeah, so Omnigen is what we call a meta harness. It's a layer that runs on top of existing agents. It can run over coding agents like Cloud Code, Codex. It can also run over custom agents and it gives you some common features across them like collaboration or security or the ability to swap between them or combine them. And it really came out of seeing what engineers were doing with coding agents here. All the really hardcore engineers had many agents running at once and many ways to orchestrate them. And we realized this is sort of a new layer that's going to emerge on top of these.
That's awesome. And what was the biggest problem managing all the agents? Was it the security? Was it the governance? Was it kind of copying and pasting instructions here and guardrails here over there? Like what was the kind of the biggest challenge you ran into internally?
Yeah, so I think with coding, a lot of engineers said, 'Well, I can vibe code my own agent orchestrator. Why do I need anything?' But the part where they all stopped short and said, 'Okay, I don't want to vibe code that.' tended to be collaboration and also security. So, collaboration, right? You got a great setup for yourself, but when you want to hand off your work to someone else, and if you or whatever they don't know how to use your thing. They have to start again. And to build a collaborative server, you need someone to host it. It needs to be secure. You don't want to just have everyone build their own. And then there's also security policies which again, you don't want to do that on your own and maybe get it wrong. So these are things that led to like yeah, maybe we should have a centralized thing. We also have a centralized like MCP and similar industry in the company and it's nice that you get dots through this as well. But these are the big ones.
Yeah, I have to imagine as agents proliferate and everyone's building more of them, having the single place to manage the governance and the auditability and the control and the security and all of that will be increasingly important especially for larger and larger enterprises.
Yeah, even for individual developers like one of the simplest things this lets me do that no existing coding agent does is I can tell it to launch a sub agent and set a budget on it. Launch an agent to read the logs and investigate this but tell me if it needs to spend more than $20. And those are usually when I had some surprise bill, it's because an agent did something stupid like trying to read millions of lines. And since OmniAgent knows about cost and supports it, I can ask it to launch any type of agent. It can very easily do that. But none of the coding agents out there actually let you do that.
Well, one of the other big advantages of using a meta harness is composability both across harnesses and across models. One of the, we talked about coding agents but one of the other examples used in the blog post is how Harvey uses both open source models as well as frontier experts together to create the best experience for their users. Are there other patterns you're seeing for composability and how is that evolving?
11 more exchanges in this transcript
Sign in free to read the rest of this interview. No card required.
Zaharia, M. (2026, July 20). Can one system run all your AI agents? | First Pass Ep. 4 with Matei Zaharia (Databricks) [Interview transcript]. Altimeter Capital. CEOInterviews.AI. https://ceointerviews.ai/interview/1103552/
MLA
Matei Zaharia. "Can one system run all your AI agents? | First Pass Ep. 4 with Matei Zaharia (Databricks)." Altimeter Capital, 20 Jul. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/1103552/.
BibTeX
@misc{zaharia2026_1103552,
author = {Matei Zaharia},
title = {Can one system run all your AI agents? | First Pass Ep. 4 with Matei Zaharia (Databricks)},
howpublished = {Interview transcript, Altimeter Capital. CEOInterviews.AI},
year = {2026},
month = {jul},
url = {https://ceointerviews.ai/interview/1103552/},
note = {Speaker-attributed transcript with timestamps}
}
