Mezmo AURA: an open source SRE agent

Andre Elizondo is Head of Product at Mezmo. Luca Berton talked to him at AGNTCon + MCPCon North America 2026 in San Jose on 23 October 2026, at the Mezmo booth, a little over an hour after AURA, Mezmo’s open source project, was named for the Agentic AI Foundation’s Open Agentic AI Innovation award.

Fortune tellers make bad SREs

The booth’s theme was “fortune tellers make bad SREs”, and Andre explained it. Mezmo is about “enabling teams to build agents to help them run production in a really safe way”. At scale, “there’s a lot of one-way doors in production”, things that can go wrong, and a fortune teller guesses: “you don’t want to be guessing in production.”

Hours, not months

Making agents ready for production, and geared towards production use cases, “is not a super simple task”. Anyone who has built agents knows how many things go wrong, and how much gets planned over time as the agents hit different walls. So in February Mezmo open sourced AURA, to shorten the loop: from wanting an agent to help manage production to having one there becomes “a matter of, you know, a couple hours more so than a couple of months or a couple of weeks even”.

AURA takes “a very Kubernetes style approach”. You define the declarative outcomes for your agent, and the system is responsible for the agent orchestration, for tracing out of the box, and for human in the loop and governance, orchestrated and instrumented from the start: “really just removing all the glue code of agents”.

Built by an SRE team

The project was started by Mezmo’s own SRE team, and many people at the company come from SRE. What went in was what they wished had existed when they built it: better root cause analysis, an easy view of the state of the systems, what led from point A to point B and why a system is failing. Then, as soon as an agent is to take a remediation, a very strict, predictable human in the loop, with an auditable trail and approvals tracked over time, “so you don’t accidentally delete production in 9 seconds”.

Two commands

AURA is Apache 2 licensed, at github.com/mezmo/aura. The readme starts with a one-line install, and an init command generates an SRE agent. It is deliberately focused on SRE workflows. Mezmo did not want to hand people “a bunch of really cool production Lego blocks” and leave them asking how to build the Death Star: “We just want to give you the Death Star out of the box.” Two commands, and AURA runs in your environment.

What sets it apart from other incident management tools, he says, is the declarative approach. Setting up an SRE agent otherwise means building “28 different things”. With AURA you maintain a simple config file, kept in source control, so you know what the agent will do at any point, and the tracing means that if the agent has a problem, it is “not going to just like hallucinate silently”. People do use Claude or Codex to manage production, and that works; what you would normally build around them to keep them safe, AURA does on day one.

From incidents to cost planning

Incident root cause analysis is the most common use. Whatever you build, systems “fail in interesting ways”. But operations takes many forms, and AURA is also used to troubleshoot security incidents and for SecOps workflows, and for performance, capacity planning and cost planning. That covers the ad hoc questions to a central SRE agent, the reactive work of incident management, and the proactive work of keeping systems optimised.

To start, there is the repository, and mezmo.com/aura explains the use cases and how to get going quickly. For anyone with an idea who doesn’t know where to start, Mezmo’s engineers will help.

More

The episode with Andre Elizondo is being produced and goes out on YouTube first. The day it was recorded: AGNTCon + MCPCon North America 2026: day two.