Git-backed AI agents survive crashes, Claude breaks editing, EKS 7-day rollback

Gastown is an open‑source framework that lets multiple AI coding agents, Claude Code, GitHub Copilot, Gemini and others, work side‑by‑side without losing context. It stores each agent's state in git‑backed worktrees, so crashes or restarts never erase progress, enabling reliable, scalable multi‑agent pipelines.

The open‑source "planning-with-files" stores task plans, findings, and progress in markdown on disk, letting agents survive context loss, restarts, or crashes. New v3 adds optional autonomous and gated modes that pause agents until the plan completes, enabling crash‑proof workflows for AI‑driven development pipelines.

The latest Claude Opus 4.8 and Sonnet 5 models often inject invented fields into the nested edits array when invoking Pi’s edit tool, causing schema validation failures despite generating correct edits. Armin Ronacher shows this regression is getting worse compared to older Anthropic models, exposing a growing misalignment between model outputs and tool interfaces.

Amazon EKS now lets you roll back a Kubernetes control plane upgrade within seven days, avoiding costly cluster rebuilds. The feature supports one‑minor‑version reversions and includes readiness checks, giving platform teams a reliable safety net for faster, more confident upgrade cycles.

Grafana Assistant Investigations now auto‑detects incidents and runs AI‑driven root‑cause analysis, then can trigger remediation actions like opening PRs or issues. The preview lets teams close the loop on alerts without manual triage, scaling "human‑in‑the‑loop" remediation across any telemetry stack.

Netflix’s TimeSeries platform now detects Cassandra partitions that exceed a byte‑threshold and automatically splits them, routing reads to the new child partitions. The metadata‑driven approach collapses read latency from seconds to single‑digit milliseconds on petabyte‑scale workloads, improving user‑facing latency and protecting clusters from GC spikes.

Grafana Cloud is consolidating alerting onto its native Grafana-managed engine, retiring the Mimir‑based data‑source alerts used by Kubernetes Monitoring. The pre‑configured rules (CPU throttling, crash‑looping pods, node failures) will now fire only under Grafana-managed alerts, so users must migrate notification channels and policies to the new system.

Express mode finishes a stack as soon as resources are configured, skipping the long stabilization phase. That can shrink a typical 64‑second SQS queue creation to under 10 seconds and a multi‑minute Lambda deletion to seconds, giving developers and AI‑assisted tools near‑real‑time feedback.