TechVoices

VIDEO PODCAST

CoreWeave’s Phil Gurbacki on AI Agents, Guardrails, and the Future of AI Development

James Maguire

October 25, 2026

As AI agents become more sophisticated, enterprises face an increasingly difficult challenge: how to ensure that autonomous systems deliver reliable results while maintaining appropriate oversight.

In this TechVoices interview at CoreWeave’s Fully Connected conference in San Francisco, Phil Gurbacki, VP of Product, Weights & Biases at CoreWeave, explains why AI agents frequently take unexpected paths, how evaluations and observability can improve their performance, and how CoreWeave’s Aria assistant helps developers build AI applications.

Gurbacki also discusses CoreWeave’s evolution beyond GPU infrastructure into a broader AI development platform, enabling enterprises to create, evaluate, and improve AI models and agents using their proprietary data.

Core Takeaways

AI Agents Discover Rather Than Know Answers

AI agents don't necessarily know the correct answer in advance. They explore possible solutions through repeated iterations, sometimes taking inefficient paths or failing to reach the desired outcome. This makes managing agent behavior increasingly important as tasks become more complex.

Guardrails and Evaluations Are Essential

As AI agents become more autonomous and operate for longer periods, organizations must establish clear success criteria and evaluation frameworks. These systems can guide agent behavior without requiring completely deterministic execution or constant human intervention.

Agent Observability Drives Continuous Improvement

CoreWeave's approach to observability extends beyond traditional monitoring and alerts. By turning agent performance data into evaluation datasets, training data, and reinforcement learning inputs, enterprises can continuously improve their AI models and applications.

CoreWeave Moves Beyond GPU Infrastructure

Through Weights & Biases, Aria, Agent Lens, and other development tools, CoreWeave is expanding beyond GPU hosting into AI application development. Its goal is to provide an integrated environment where enterprises can build, evaluate, and operate AI systems while retaining control over their AI development and outcomes.

KEY QUOTES

AI Agents Don't Know the Answer — They Discover It

“I think a lot of folks think that agents know the answer. They believe that agents have a level of intelligence where they can generate the answer, they can give you the exact answer. And that's actually not the case. The agents discover the answer. And so they go through a number of different iterations and they try one approach, they walk it back, they try another approach, they walk it back and they continue to iterate until they find the answer, which in some cases could be quite expensive if they take a suboptimal path.

“And in some cases it could end up looking like a lot of misdirections or missteps or they can veer off the ideal path. And so if you have in your mental mind the ideal trajectory, an agent might veer off of that. They might make it back and they might not.”

Why Evaluations Are Critical for Autonomous AI Agents

“One piece is, as we discussed earlier, the agent is exploring a potential solution space. And so you want to have guardrails to make sure that you provide a container and guardrails and a space for the agent to explore. I think the second thing to think about is what does success look like?

“And as the agents become more intelligent, as the agents become more autonomous, as they're running for longer and longer periods of time, the actual hard problem is defining what success looks like and defining what the ultimate outcome is. Evaluations are a really great way to do that. And so if you have good evaluation criteria for the agents that you're building, you're able to then steer them in the right direction and make sure that they're performing on the dimensions that you care about.”

Agent Observability Must Go Beyond Traditional Monitoring

It's important: "the ability to take observability data and turn it into something useful from a training perspective or from an evaluation perspective. And that's totally different than observability that you might have seen in the past. Traditional machine learning observability was all about setting alert conditions and monitoring distribution of data.

“It's actually totally different for agentic systems. You need a new way of evaluating the results that are coming back. Of course you're monitoring and alerting, but there's another layer of intelligence that's required to generate the insights, to generate the data sets and ultimately build the next iteration of the model and agent solution that you're building.”

AI Development Requires a New Generation of Tools

“Part of our thesis is that the ability to do AI development, whether it's AI powered application development or agent development, it's a fundamentally different process than software development. So many of our customers are very familiar with software development. They understand that workflow. It's been hardened over decades of investment and work to build processes.

“Agent and AI development require a new set of tools and a new set of services to help researchers and developers be successful building production grade systems. And so that's where we're focused: making sure that the developers that are developing the technology, that the researchers that are building new models and researching new insights, they have the tools they need to be successful building. And I think most importantly, we want our customers to own their AI and own their outcome.”

ABOUT THE AUTHOR

James Maguire

Executive Director

An award-winning journalist, James has held top editorial roles in several leading technology publications, covering enterprise tech trends in cloud computing, AI, data analytics, cybersecurity and more. He regularly communicates with industry analysts and experts and has interviewed hundreds of technology executives. James is the Executive Director of TechVoices.

Cloudera's Venkat Rajaji: advantages of open source AI

Snowflake's Christian Kleinerman: AI and the speed of data access

David Linthicum: LLM trends

Extreme Networks' Nabil Bukhari: how AI democratizes technology

Komprise's Krishna Subramanian: an AI model for every task

Ahead of the curve, every week.

Get in touch with us.

Ahead of the curve, 
every week.