DevOps · Sunday, 23 August 2026 · Issue 81 · 7 stories

GitHub outages, OVH price spike, and agent density

You've seen the individual RCA for each GitHub outage, but what if the real root cause is one you never read? And while OVHcloud is hiking prices 87%, Google is showing how to run 250 agents on 8 pods, two sides of the same AI coin.

Containers & Orchestration

Hibernate agents at scale: 250 stateful actors multiplexed onto 8 Kubernetes pods

Google has open-sourced Agent Substrate, a low-opinion control plane for running AI agents at high density. It exploits the fact that agents are mostly idle, snapshotting state in volatile RAM so an actor can suspend and resume in under a second onto any free worker. That buys 30x oversubscription, a demo juggles 250 stateful actors across just 8 pods. It runs atop Kubernetes and gVisor, so it hosts agents built on any stack.

github.com ↗7 min read at source

Observability & Reliability

13 GitHub outages in 17 days, and still no collective root cause

GitHub logged its 13th separate incident in the first 17 days of August, each with its own individual RCA. Treated singly, they read as isolated blips; strung together, two recurring failure modes emerge, capacity headroom and change management, that nobody at GitHub has synthesized. That gap is the real story for reliability teams: per-outage postmortems keep missing the systemic pattern underneath.

cyberkendra.com ↗12 min read at source

Cloud & Platform Engineering

Cloudflare ships Kitesurf, a browser for AI agents that drops Chromium entirely

Cloudflare built Kitesurf, an agent-first browser engine that runs as Rust-compiled WebAssembly in isolated Workers, and throws out everything a human browser needs that an AI doesn't. No tabs, extensions, or pixel-perfect rendering; just structured extraction, screenshots, and 3-7x less CPU and memory than Chromium. It speaks Chrome DevTools Protocol so Puppeteer and Playwright scripts run unchanged, and Cloudflare plans to open-source it. The catch: no video, WebGL, or long-lived authenticated sessions yet.

blog.cloudflare.com ↗17 min read at source

AI RAM squeeze reprices non-AI clouds: OVHcloud hikes servers up to 87%

AI's hunger for high-bandwidth RAM is collateral-damaging the servers that run everything else. OVHcloud will raise most of its catalog from September, 87% on 2026 gaming boxes, 40-59% on recent servers, after its memory costs hit 6x the June 2025 level, heading to 9x by September and 12x in early 2027. Klaba expects the squeeze to last until 2028. Tellingly, AWS, which designs its own accelerators, has barely moved.

infoq.com ↗5 min read at source

Sovereignty moves into the cloud architecture: control plane is the new boundary

Sovereignty has left the policy debate and landed in the architecture. DORA, NIS2 and the EU Data Act mean "eu-central-1" is no longer enough: teams now have to answer who controls the environment and where its logs, policy engine and control plane live. The question shifts from data residency to data sovereignty, who actually supplies and governs the stack.

versa-networks.com ↗4 min read at source

DevSecOps

One DRAM register bit gives away AMD's PSP, SMM, and microcode

Christopher Domas' skitter-creek-bath-salts flips translation bits in AMD Family 16h DRAM controllers, scrambling physical address maps so unprivileged code lands inside memory carveouts the kernel can't even enumerate. Upstream fences, hypervisor EPT, SMM TSEG bounds, PSP carveouts, validate only the untranslated address, so the swizzle below slips through. That unlocks PSP firmware, SMM RAM, and microcode buffers. No CVE, no patch: the registers simply can't be locked, and newer AMD generations just omit them from the datasheets.

github.com ↗31 min read at source

Debian tables a vote to ban LLM-authored code contributions outright

Debian's committee has tabled a general resolution that would ban direct LLM-assisted contributions to Debian code, docs, and translations, styled as a defense of what makes Debian Debian. It's actually an eight-way vote with counterproposals from conditional allowance to an outright "climate destruction" rejection. Upstream patches and packaging of third-party AI-coded tools stay exempt. The real question, one observer notes, is how big the provenance sticker should be.

thenewstack.io ↗6 min read at source

You’re caught up.

That was all 7 stories in the 23 August issue.