Humans of Reliability

What happens when AI agents become direct users of your infrastructure?

Michael Kutsch, Staff SRE and Team Lead for Cloud Foundations at PostHog, is building toward an agent-native platform. His team gives agents an index to focused infrastructure context, then packages repeatable work as skills that can call deterministic scripts when reliability and governance matter.

Michael joins Sylvain Kalache to explain why this technical model depends on genuine engineering autonomy. They discuss how PostHog's product teams contribute directly to infrastructure, how layered documentation improved agent-generated pull requests, where agents help during incidents, and why the enduring work of infrastructure engineering is moving toward architecture, judgment, and review.

AI is rewriting the reliability-engineering career path from the first application onward. James Duffy, Principal SRE at Fabric, explains why applications are becoming harder to distinguish and why interviews must stop rewarding output and start revealing production judgment.


He explains how he replaced generation-focused exercises with a review of deliberately flawed AI code and how job seekers can use AI for skeptical feedback without surrendering their own voice.

The interview follows that path beyond hiring. James argues that companies abandoning junior roles are creating a future talent gap, and that continuous learning is the most durable skill for an SRE career in which the number of technologies and responsibilities keeps growing.

The race to build capable AI agents often focuses on model intelligence. Anthony Alcaraz argues that the harder, more durable problem is context: giving agents the right information, relationships, tools, and constraints at the moment they need them.

Anthony is a Senior AI/ML Portfolio Growth Manager at AWS and co-author of O'Reilly's Agentic GraphRAG. He explains how graph-based architectures can improve retrieval, memory, planning, and multi-step reasoning while capturing the tribal knowledge that conventional enterprise systems leave behind.

The conversation also looks beyond architecture. Anthony describes self-improving agent systems, feedback loops that connect technical performance to business outcomes, and a future where small teams manage large fleets of agents without losing sight of customers or value.

AI is giving attackers the ability to scale phishing, discover vulnerabilities, and build increasingly autonomous malware. But it is also helping defenders investigate alerts, eliminate noise, and respond before threats become incidents.

Nir Soudry, Head of R&D at 7AI, joins Humans of Reliability to explore the rise of agentic cybersecurity. He explains why humans must remain accountable for AI decisions, how security teams can introduce agents gradually, and why adopting AI should be treated as an ongoing partnership rather than a simple software installation.

We also discuss AI-generated code and its impact on reliability, 7AI’s claim that automated investigations can remove roughly 90% of alert noise, and the industry’s evolution from reactive security toward proactive, policy-driven prevention.

Will generative AI create more incidents—or help us stop them before they happen?

When incidents pile up fast enough, every part of the company bleeds: support is fielding angry customers, AEs are on apology calls, and engineering is burning cycles on retrospectives instead of shipping.

For Eran Kampf (VP of Engineering at Twingate) where the product is the network, that was the moment he made a call most engineering leaders won't: stop all feature work for a quarter and fix reliability.

In this episode, Eran walks through what lead to that decision, why best-in-class engineering practices (full test coverage, Terraform provisioning, comprehensive monitoring) turned out to be just the baseline for four nines and what it took to rebuild trust internally and with customers. 

"Hey Claude, where's my database?" The answer came back: "Oh, sorry, your database is gone." Alexey Grigorev, founder of DataTalks.Club and creator of the Zoomcamp courses that have taught data and ML engineering to over 100,000 people, knows Terraform well enough to cover it in his curriculum. That didn't stop a chain of small, reasonable-sounding decisions from ending in an AI agent running terraform destroy against his live production database. 

Every pilot who's never had an engine failure is still ready for one. The same can't be said for most software engineers facing their first major incident. 

Hamed Silatani, co-founder and CEO of Uptime Labs, and former Head of Reliability Engineering at IG Group, has spent two decades watching engineers learn incident response the hard way: alone, under pressure, with no training. 

A moment in a hospital delivery room, where a team of nurses responded to a crisis with clockwork precision, crystallized what he'd long suspected: IT spends heavily on tooling and process but almost nothing on training the humans who have to use them under fire. 

In this conversation, Hamed breaks down the three skill categories that matter in incident response, explains why communication is always the first pain point leaders name, and makes the case that AI-driven complexity is about to make human incident skills more important, not less.

For nine years, Maria Vechtomova was shouting about monitoring. Nobody cared, until LLMs arrived. 

As co-founder of Cauchy, Databricks MVP, and one of the most followed voices in MLOps, Maria has watched the field evolve from hand-built experiment trackers to today's flood of observability tools, and her central claim might surprise you: globally, nothing has changed. 

The fundamentals are the same: track your code, data, and models so you can roll back when something breaks. 

What did change is the surface area. Tools, prompts, embeddings, agents, every component shifts behavior unpredictably, and business metrics often become the only signal left. Maria gets into why most teams still can't roll back cleanly, why 40% of ML projects ship with no monitoring at all, and why she believes the next era of MLOps will be its biggest yet.

Most incident response advice focuses on tools, alerts, and post-mortems. Gandhi Mathi Nathan Kumar, Principal Incident Commander at Twilio, with 14 years running calls that have pulled in up to 100 responders, argues the work that actually matters happens in the first 15 minutes. 

In this episode, Gandhi walks through what he calls the golden hour: the window where you decide whether you know what's broken, who belongs on the call, and whether to chase the root cause or reach for redundancy. 

He gets into why mitigation has to come before diagnosis, why customers trust your status page more than your engineers, and why he once sat with a stopwatch counting how many clicks it took to declare an incident. Along the way: the human side leaders keep underinvesting in, the math of on-call fatigue, and where AI is actually pulling weight in the incident commander seat.

A centralized SRE team of 600 engineers as the first line of defense for every incident works - until the business asks you to spread that responsibility across 6,000. 

Cliff Snyder, senior SRE at Multimedia and a decade-long veteran of LinkedIn's SRE org, walks through the 18-month project to democratize incident response: replacing a patchwork of Jira, Google Docs, and in-house tooling with a purpose-built platform designed for engineers who may have never been on-call before. 

He gets into the tradeoffs that shaped the rollout — flexibility versus opinionation, exec pushback on even renaming "postmortem" to "retrospective," the reporting miss that almost pushed launch back, and why MTTR is the wrong number when a single incident can burn a thousand engineer-hours. An honest look at what it actually takes to move incident management from tribal knowledge to federated practice at scale.