🔥 Top Story
Spotlight on SIG Storage
Kubernetes SIG Storage owns the core storage primitives (PersistentVolumes, PVCs, StorageClasses) and the out-of-tree CSI driver model that replaced in-tree plugins. Recent work includes Volume Group Snapshots reaching GA in v1.36, Changed Block Tracking for efficient incremental backups, and COSI for object storage. The group is also adapting to emerging AI workloads and their new storage demands.
💡 Why it matters: Maturing CSI drivers plus Volume Group Snapshots and CBT mean stateful-workload backup and DR can now be built on vendor-neutral standards. Cluster operators should plan migration away from in-tree plugins toward CSI-based snapshot and resize pipelines.
🔗 Read more · Kubernetes
DevOps & Infrastructure
Xiaomi’s MiMo Code claims it beats Claude Code past 200 steps
The New Stack
Xiaomi open-sourced MiMo Code, a terminal-native coding-agent harness it claims beats Anthropic's Claude Code on agentic tasks running past 200 steps—though the numbers are self-reported from its own beta and a 576-developer survey. The piece argues the real battleground is the 'endurance gap'—how far an agent sustains a single objective across hundreds of dependent steps—citing UC Berkeley's code-graded 'Agents' Last Exam' where even the strongest setup (Codex + GPT-5.5) scored under 50% on the easiest tier and under 10% on the hardest.
💡 For engineers wiring agents into CI/CD or ops pipelines, this is a warning to measure long-horizon endurance as its own evaluation axis, not demo performance. Make 'can the harness externalize state and resume from a checkpoint' (e.g., Arbor's hypothesis tree, Claude's 5-level subagents) a procurement criterion, so you catch agents that silently collapse around step 30 and hand back plausible-but-wrong artifacts.
What your logs can’t tell you when an AI agent acts alone
The New Stack
The article distinguishes a surface log ('something happened') from a real audit trail that captures who, what, when, where, and before/after state—and stresses that when an AI agent acts autonomously, the audit trail is the only source of truth because no ambient human context exists. It argues you must log agent identity, the authorization chain, and scope boundaries, driven by SEC/NIS2 rules, SOC 2, and Verizon's 2026 DBIR (vulnerability exploitation now the #1 initial-access vector at 31%). Well-structured, immutable, queryable logging has shifted from an internal asset to a revenue and procurement differentiator (illustrated by Webflow's in-product activity log with AI-vs-human attribution).
💡 In an era where agents provision resources, change settings, and delete data, logging design must treat agent identity, authorization chain, and scope as first-class fields, and ensure logs are immutable, queryable, and retained long enough (6+ months, not just 30 hot days). This is not mere compliance but a concrete competitive edge that clears security reviews and vendor evaluations—so it belongs on the SRE/platform team's priority list.
PagerDuty’s CAIO says most AI incident tools are missing a critical layer
The New Stack
PagerDuty's CAIO argues the critical layer most AI incident tools are missing is the agent 'harness.' Bolting on MCP connectors isn't enough—agents need access to the right data (code changes, logs, metrics, traces, runbooks, service topology, on-call info), short- and long-term memory, and the judgment to know which actions are safe. For trust, the harness must also provide transparency and control: configurable allowed/forbidden actions, human-approval gates, and inheritance of team permissions.
💡 SRE/DevOps teams adopting AI for incident response should invest in harness design—injecting the right operational context, a memory layer that updates and invalidates facts as hypotheses shift, and action permissions with approval gates—rather than the model or MCP wiring alone. A practical starting point is using historical incident data for pre-production risk scoring of code changes, as the article describes.
This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.