Operations | Monitoring | ITSM | DevOps | Cloud

"Hello world, this is your agent speaking!" The agent loop! The LLM is calling tools, the answers are sensible, and the sky's the limit. Now, as you look forward to production, you look for a composable toolset, something that can grow with your use case and system needs. That's what we created with Honeycomb Canvas: a collaborative investigation space where AI agents help you understand, fix, and learn about your system.

As of today, I’ve drafted this post upwards of 10 times – it’s old enough that the version I first started working on was called “Reflections on 1 Year of SRE Management” (I’m currently at 2.5 years). But everything I learned during that first year became critical for the next.

At O11yCon, we talked to engineering teams across the industry, and the numbers are starting to get genuinely wild: Mixpanel DevOps Engineer Eddie Bracho told us their engineering team is generating 50% more PRs than before AI came into the mix (sorry). That kind of velocity is exciting, but it's also a pressure test for every part of your stack that isn't writing code, including your observability practice. Here's what we're hearing from customers about how that's playing out.

AI agents are nondeterministic, multi-step, and opaque. When one fails in production, "the model said something weird" is the cheapest, most useless line in your incident postmortem. To debug agents the way they actually run, you need telemetry that captures all of it, in order, with enough context to reconstruct what happened. The OpenTelemetry GenAI Semantic Conventions give you a vendor-neutral way to do exactly that.

IT’S HERE it’s here it’s here it’s here!!!! The second edition of Observability Engineering is available for download, and since Honeycomb is the sponsor, you can now download it from our website (the dead tree version will take another month). This is a strange time to be writing a book.

A few weeks ago I wrote about a customer’s refund request that stopped halfway through at 11:47 p.m. on a Tuesday night. That post walked through the 40 minutes it took to work out what happened when an agentic application had a problem: a tool retried against a rate-limited payments API, the error responses filled up the context window, and the agent gave up. The whole reason we built Agent Timeline was to turn that 40 minutes into five. To reduce MTTR. To solve the problem and get back to sleep.

Observability has always been important, and much like any core capability in your business, the value needs to be understood. For years, the value of observability was predictable. It was uptime, error rates, MTTR, and likely tool consolidation. That was enough to be able to show progress. These are foundational, tablestakes metrics—and they still matter, but they aren’t enough.

This is the fourth installment in the Graviton retrospective series we've been writing since 2021. The methodology is the same one I always reach for: hold the workload constant, run both generations on the same Kubernetes namespace concurrently, and let the per-pod numbers speak.

When a measure becomes a target, it ceases to be a good measure. Charles Goodhart, 1975 You’ve probably read this quote in relation to any number of things over the years. People complaining about arbitrary metrics like PRs merged, lines of code produced, and now, token usage. But is the era of tokenmaxxing over before it even began? The rise of token leaderboards to the death of token leaderboards at companies like Amazon seem to have taken place in less than three months!

The OpenTelemetry Collector is usually deployed as a long-running process: a sidecar, a DaemonSet, an EC2 instance, a docker container on my computer. It sits there listening for telemetry. That's fine when I want to send telemetry all day, but not when telemetry is rare. Like right now, when I have an agent defined on AgentCore, and it runs a few times a week maybe. Or my website that hardly sees any traffic. Can I run the OpenTelemetry Collector as a Lambda function?