Deep dives into Site Reliability Engineering, AI in production, and scaling infrastructure gracefully. Written by the team building the future of SRE.
A Kubernetes outage rarely announces the deployment that caused it. Start from the alert, prove the change window, identify the ReplicaSet and image digest, then trace the rollout back to its CI run and commit.
Two customer investigations show how a context graph was key to helping agents retrieve the right data by tracing changes and shared dependencies.
We were at SREday London on September 24 to talk cascading cloud failures. Photos from the day, plus a walkthrough of the talk: replaying the Cloudflare outage of 18 November 2025 with a dependency graph in hand.
During a production investigation, an agent must connect applications, machines and alerts. Across five integrations and ten tasks, Anyshift graph queries used 83% fewer model tokens and took 47% less time on average than native MCPs.
We rebuilt our agent's architecture to separate operational knowledge, model runtime, and side effects. When we added OpenAI Codex support, the pull request only touched one file
Unified production context makes AI agents cheaper, faster, and easier to scale. In our live benchmark, resolving relationships before inference cut model-token usage by 83.5%, while reducing response time and estimated cost without changing the task outcome.
We tested six models on a bounded task: turn a plain-English infrastructure question into one supported query. Deterministic normalization and execution erased most output differences, leaving Haiku and Luna slightly ahead of Opus on final-answer accuracy. Luna then matched Haiku at 21% of the observed cost.
Annie (our AI SRE agent) had institutional memory from ACE, the agentic context-engineering loop that curates cheatsheets from past runs. It worked, but clients kept catching her trusting stale entries or missing answers buried in her own bloated context. So we added five things on top: (1) a fixed set of memory items always presented to the agent, (2) per-query retrieval over the rest of the memory store, (3) an agent-optimized index of that store, (4) the ability for the agent to query the store mid-run, and (5) tried-and-true memory freshness mechanisms. Production context, now optimized by the AI using it. Here's the reasoning and what a few weeks in production say.
An AI agent in the incident channel can run kubectl and read a dashboard. What it can't do is judge whether the last deploy is the suspect or a red herring. We open-sourced the SRE skills that encode that judgment, runnable offline against fixtures with no credentials.
Private equity firms are pushing AI across portfolios, but the hard part is no longer experimentation. It is turning sponsor-level AI pressure into governed, production-aware operating workflows that actually move EBITDA.
Burned 25 minutes on a Friday-morning page before I realized the responsible commit was in another team's repo. This is the four-command sequence I now run when an alert lands and `git log` on my own service comes up empty, with the outputs at each step and where the search space gets cut.
How AI agents learn your infrastructure. A walkthrough of the DevOpsCon Amsterdam 2026 talk: the gap between LLMs and production, structural context as a versioned graph, and ACE for self-improvement without labels.
Temporal workflows stuck in Running with zero pollers, and Temporal still reports a healthy task queue. The root cause lives one layer down: a CrashLoopBackOff in the Kubernetes worker pod, caused by a single bad environment variable. A walkthrough of debugging Temporal workers on Kubernetes the manual way (10 minutes), then with an infrastructure context layer that bridges the two systems (seconds).
The 10 best AI SRE tools in 2026 compared by architecture, root cause analysis, remediation, and change awareness — from Anyshift's versioned graph to Resolve AI's autonomous agents.
AI agents start every run from scratch. ACE (Agentic Context Engineering) gives them institutional memory that evolves through use, cutting root cause analysis time by 30%.
How Anyshift chose Neo4j for building a temporal infrastructure knowledge graph and lessons learned over a year of production use.
The limits of telemetry-only AI approaches to SRE and why topology is the missing piece.
Key takeaways from the Civo Navigate SF conference.
A talk recording from DevOpsDays Dallas about the challenges of debugging infrastructure quickly.
![[Featured in Tessl] DevOps with AI: Identifying the impact zone, with Roxane Fischer](/images/blog/devops-with-ai-impact-zone.png?dpl=dpl_6bJJjhavN8iW2QqMumhhJpz3y9CQ)
A featured interview on Tessl about DevOps with AI and identifying the impact zone.
GenAI is everywhere. But very often, the cool and exciting demos don't work the same way in production.
Managing infrastructure at scale is a complex endeavor that demands meticulous planning, robust tooling, and continuous adaptation.
Most infrastructure debugging sessions blow past the one-hour mark for the same five structural reasons: scattered visibility across cloud accounts, missing historical state, terraform plan output that hides downstream impact, runbooks that lag the live infrastructure, and post-merger environments that no one has fully mapped. A walkthrough of each, with concrete examples and what reduces the time.
Three structural patterns recur in growing infrastructure orgs: single-repo bottlenecks where dozens of teams share one approval queue, ClickOps and dead IaC code that drift outside any state file, and module version fragmentation that quietly bypasses security patches. A walkthrough of each, with the practices that contain the blast radius.
Get a monthly digest of our best engineering articles, SRE case studies, and Anyshift product updates. No spam, just signal.
