Mezmo CEO Tucker Callaway on the Future of AI Agents, Observability and Autonomous IT Operations

Mezmo's AURA, an open-source SRE agent harness designed to help teams build and deploy production-ready agents within their existing environments, was recently named AI-powered Enterprise Agent Solution of the Year by the 2026 AI Tech Ascension Awards. We spoke with Tucker Callaway, CEO of Mezmo, to discuss how AI agents are reshaping observability, the challenges of moving from AI demonstrations to production deployments, and why governance, transparency and trust will be critical to the future of autonomous IT operations.

How do you think the rise of AI agents will impact observability?

I think observability is headed for a pretty fundamental reset. Historically, observability has been built to help humans understand systems through dashboards, queries, and investigation. As agents take on more operational work, that model changes. Agents need useful context delivered directly into the workflow. Agents don’t work that way. They need useful context delivered directly into the workflow.

That means the value starts to shift. Storing massive amounts of data and giving people better ways to search it becomes less important. Turning noisy telemetry into focused evidence becomes much more important. Agents don’t need more data. They need the right data.

We’ll also need to observe the agents themselves. What did the agent see? Why did it make that decision? What tools did it use? Did it stay within the boundaries we set? Observability becomes as much about establishing trust in autonomous workflows as monitoring the underlying infrastructure. ⁠

There’s a lot of energy around AI SRE right now. How can companies separate reality from hype?

The demo is easy. Production is hard.

You can connect a model to a few tools and generate a pretty impressive investigation. The real question is whether it works reliably in your environment, with your data, when something important is actually failing.

I’d look at practical outcomes. Did it reach a correct, validated conclusion faster? Can the SRE interrogate the result and see the evidence? Can it reuse evidence and outcomes from previous incidents? Can the team control what it’s allowed to access and what it’s allowed to do?

Companies also shouldn’t have to choose between a passive copilot and an agent with unrestricted access. Start with investigation. Measure the results. Require approval for consequential actions. Give the agent more authority as it proves itself. That’s how you turn an interesting demo into something your on-call team will actually trust. ⁠

⁠ ⁠

⁠Do you think enterprises feel pressure to adopt AI solutions? Are there concerns they should be mindful of about moving too fast?

There’s definitely pressure. Every executive team is asking how quickly the company can take advantage of AI, and development teams are already shipping faster because of it.

The concern is that operational maturity may not keep pace. More code means more releases, services, dependencies, and change entering production. Companies can end up creating more operational complexity than their teams can reasonably understand or manage.

I don’t think the answer is to wait. Start with a bounded workflow and be clear about what success looks like. Decide what the agent can access, what requires approval, and how its decisions will be recorded. Plan for what happens when it’s wrong, because eventually it will be.

The simplest way to put it is: humans set the boundaries, agents do the work. Autonomy is where this is going, but agents have to earn trust along the way. ⁠

⁠ ⁠

⁠You recently announced an open-source project called AURA for building production-ready agents. Can you explain more about the project and its benefits?

AURA is our open-source SRE agent harness. It runs in the customer’s infrastructure and coordinates workflows across the observability tools, runbooks, ticketing systems, collaboration platforms, and models they already use.

We built it because generic agent frameworks leave a lot of production work to the customer. SRE teams need to inspect how the agent reached a conclusion, control the systems it can access, and adapt workflows to their own environment.

Open source matters when an agent is working in production. Teams need to see how it operates and understand why it made a decision. I don’t think companies will be comfortable putting a black box in the middle of incident response.

AURA handles the workflow. The Mezmo platform provides the governance, auditability, memory, and operational context around it. Teams can begin with investigations and gradually expand the agent’s authority as confidence grows. ⁠

⁠ ⁠

What does Mezmo see as the core challenges it hopes to help solve in 2027 when it comes to building with AI in the world of software?

The big challenge will be operating everything AI helps us build.

Software is becoming much easier to create and change. That doesn’t make it easier to run. In fact, the operational surface area is likely to grow much faster than our ability to understand it manually.

Agents will need focused context instead of unlimited access to raw data. Enterprises will need a reliable record of what each agent saw and did. They’ll also need governance that keeps people in control without asking someone to approve every individual tool call.

That changes the role of the SRE. People move up the stack and spend more time defining policy, boundaries, and success conditions. The agent handles more of the investigation and operational work within those constraints.

I don’t think the winners will simply have the smartest agent. They’ll have the best environment for putting agents to work safely: useful context, clear controls, an auditable record, and a way for each workflow to earn more authority over time.