Blog
Insights, analysis, and updates from the AI agent economy. Browse by tag · Browse the archive.
PostgreSQL 18's Async I/O, OAuth, and pg_upgrade --swap: What Your Operator Image Should Actually Turn On
PostgreSQL 18 shipped async I/O, OAuth 2.0 authentication, and a faster pg_upgrade --swap mode. A concrete recommendation for which of the three a self-hosted PaaS's CNPG or Zalando-managed Postgres image should turn on today, and which ones aren't ready yet.
PocketBase vs. a Cluster API Fleet: The Three Walls That Tell You When to Switch
PocketBase's single Go binary handles 10,000+ realtime connections on a $4 server — until sustained multi-writer contention, HA requirements, or hard tenant isolation hit walls a Cluster API fleet is built to clear.
Paketo Buildpacks Loses Its GCP Sponsor: What a Cloud Vendor Walking Away Means for Betting a Build Layer on the Open Standard
In 2025 Google pulled the sponsorship paying for Paketo's container registry hosting, giving it two weeks' notice. Here's why the project barely noticed, and the four-layer funding audit any team betting its build pipeline on someone else's open-source project should run first.
Netlify Killed the Seat Tax, Vercel Kept the Credits: Recomputing the 2026 Self-Hosting Breakeven
Netlify's April 2026 update killed per-seat billing and repriced bandwidth in the same changelog entry, quietly breaking the '800GB on a 5-seat team' rule of thumb that's still all over comparison sites. Here's the recomputed breakeven — and where a single self-hosted Hetzner box lands on the same table.
Ephemeral Preview Environments on Kubernetes: A Namespace-Per-PR Recipe for a Git-Push PaaS
A working ArgoCD ApplicationSet recipe for namespace-per-PR previews, what it actually costs in cluster resources at 10 vs 50 concurrent PRs, and the quota/network-policy/TTL guardrails that keep a busy repo from starving a shared cluster.
MCP Tool Schemas Are Eating 72% of Your Context Window: How to Design an Infrastructure MCP Server That Doesn't
A production benchmark shows MCP tool schemas can eat 72% of an agent's context window before a single query runs. Here's why, how Pinterest fixed it at scale, and how to design an infrastructure MCP server that doesn't repeat the mistake.
MCP Tasks Gets Retry Semantics and Expiry Policies: The 'Call Now, Fetch Later' Pattern for Deploys That Outlive an HTTP Timeout
MCP's Tasks primitive just went Final: client-generated task IDs make retries idempotent, keepAlive sets result expiry, and the 2026-07-28 spec reshapes both into a formal extension — what a deploy-from-chat MCP server needs to implement for deploys and rollbacks that outlast an HTTP timeout.
MCP's Stateful Session Bottleneck: Why the 2026 Transport Roadmap Is Racing to Make Agent Servers Horizontally Scalable
MCP's official spec drops protocol-level sessions on July 28, 2026 — here's why that's the fix for the exact bug that makes a deploy-from-chat MCP server a worse single point of failure than the app it manages.
MCP Server Cards Won't Advertise Your Tools — Here's What They Actually Do
A draft MCP proposal lets servers publish machine-readable metadata at a .well-known URL — but the real schema deliberately leaves out tool names, and that omission changes what publishing one actually buys a self-hosted platform.
Subscribe
New posts land in your reader as soon as they publish. Pick a format — all three carry the same posts.
Current feeds keep roughly two days of posts so daily polling does not miss a burst. Older entries stay reachable from the feed's next-page link in readers that follow it, or from the blog archive.
Following one topic instead? Browse tags