AURA, an open source SRE agent

Why this conversation

An agent that is allowed to act on production can do things that cannot be taken back, and the work of making it safe usually falls to the team that runs it. Andre Elizondo heads up product at Mezmo, whose open source project AURA is an SRE agent with much of that safety work already done. The host found him at the Mezmo booth in San Jose on Friday 23 October, the second day of AGNTCon + MCPCon North America, not long after the project had won an award. Everyone else the show met that Friday is in AGNTCon + MCPCon North America 2026: day two.

Fortune tellers and one-way doors

The booth carried a slogan, “fortune tellers make bad SREs”, and the host wanted to know why. Mezmo, Andre explains, is focused on letting teams build agents that help them run production safely. Running it at scale means facing many one-way doors, decisions that can go badly wrong, and a fortune teller is someone who guesses. As he puts it, “you don’t want to be guessing in production.”

Getting an agent ready for production use is hard, he says. Anyone who has built one has seen how much can go wrong, and how plans pile up as the agent runs into one wall after another. Mezmo open sourced AURA in February to shorten that path: from deciding you want an agent to help with production to having one running should take a couple of hours, not weeks or months.

The design borrows from Kubernetes. You describe, declaratively, the outcomes you want from the agent. The system then handles the orchestration, turns on tracing from the start, and wires in human in the loop and governance, so the glue code that agents usually need goes away.

The project was started by the SRE team inside Mezmo, and plenty of its staff come from SRE themselves. What went in was what that team had wished for: better root cause analysis, an easy picture of the state of a system, and an account of what led to a failure. When the agent wants to fix something, a person has to approve the step first, and every approval is kept as an audit trail.

Two commands, one config file

AURA is Apache 2 licensed and lives in the mezmo/aura repository on GitHub. Near the top of the readme there is a one-line install, and an init command then generates an agent built for SRE work. Andre says that narrow focus is on purpose: people should not be handed a pile of parts and left to assemble a finished system themselves. Two commands, and AURA is running in your environment.

The host pointed out that other incident management systems sound similar. For Andre the difference is the declarative approach. Without it, setting up an SRE agent can mean building 28 separate pieces. With it, there is a simple config file to look after, which can sit in source control. “You know predictably what the agent’s going to do at any point.” Because tracing is there from day one, an agent that runs into trouble cannot quietly hallucinate and do things nobody expected. Claude and Codex are great, he says, and people do use them to manage production; the point is that the safety you would otherwise have to build around them comes with AURA from the start.

Incidents, security and cost

Finding the root cause of an incident is what people use AURA for most often. “The problem that everybody has is that no matter how good you build your system, they fail in interesting ways.” Operations takes other forms too, and some teams use AURA to troubleshoot security incidents and for SecOps workflows, while others use it for performance, capacity planning and cost planning.

Andre sorts the work into three kinds: ad hoc questions and tasks for a central SRE agent, reactive incident management, and proactive work to keep systems as optimized as possible. The host was pleased that somebody was taking care of the proactive side at last.

The quickest way in is the repository. For a gentler start than a readme, mezmo.com/aura describes some of the use cases and how to get going quickly, and Mezmo’s engineers, he says, are glad to help anyone who has an idea for a use case and does not know where to begin.

When it comes out

The episode is edited and in the release order. Like every episode it goes out on YouTube first, then on Apple Podcasts, Spotify and Amazon Music.