Why your company should (try to) build its own AI SRE
The New Stack Nick Lucchesi
Chronosphere folks say build your own AI ops agent before buying one. Mapping your systems yourself is the real value, not just the tool.
Agentic coding tools have made writing software faster than ever, which sounds great until something breaks at 3 a.m. and nobody on the team actually understands the machine-generated code that failed. That's the uncomfortable spot site reliability engineers find themselves in right now. Code ships faster, but the humans tasked with fixing it when it breaks are stuck doing archaeology on systems they didn't build.
Sam Farid and Nate Heinrich, who work at Chronosphere (now part of Palo Alto Networks), think the fix is more agents, not fewer. On a recent episode of The New Stack podcast, they argue that root-cause analysis is exactly the kind of grinding, pattern-matching work AI models are getting better at, even though benchmarks like OpenRCA show there's still plenty of room to improve. Their pitch isn't subtle: fight fire with fire.
What's more interesting is the advice that follows. Heinrich says he actively encourages companies to build their own AI SRE in-house before shopping for a vendor product, even though Chronosphere sells exactly that kind of product. His reasoning is that the real value isn't the finished agent, it's the process of building one. To make an agent useful for root-cause work, a team has to document how its systems actually connect and behave, often producing something as simple as a Markdown file the agent can reference later. That documentation exercise forces institutional knowledge out of people's heads and into a format machines, and other humans, can actually use.
Of course, there's a business logic underneath the generosity. Chronosphere is betting that once a company goes through this exercise, it'll realize it needs telemetry pipelines to capture logs and traces, and an observability layer to store and correlate all of it, both things Chronosphere happens to sell alongside its own AI SRE product. So the free advice doubles as a funnel. But it's a funnel built on a genuinely useful insight: you can't hand an agent a mess and expect clarity back.
Still, the framing here matters more than the vendor angle. Neither Farid nor Heinrich pretended their tool was the only path forward, which is rarer in this space than it should be. Their real argument is that companies who wait for a perfect off-the-shelf AI SRE are skipping the hard, valuable work of understanding their own systems, work that pays off whether or not they ever buy anything from Chronosphere.
My take
I like advice that survives contact with its own sales pitch, and this mostly does. The honest bit is that mapping your systems is valuable homework regardless of vendor, and too many teams skip it hoping a tool will magically know things nobody bothered to write down. But let's not pretend Chronosphere is being purely altruistic here; they're betting the homework leads straight back to their shopping cart, which is a very normal, very fine thing for a vendor to do as long as you go in with eyes open.
Read more about this at: The New Stack
