Wed 08 Jul 26
The coding agent clocks in — auditing a province's entire codebase, mounting its own filesystem, and getting fact-checked on its own hype.
Twenty hours to read every line the province runs.
Alberta's technology ministry pointed Claude Code at its whole estate — 1,280 applications, 3,400 repositories, 466 million lines — and had it scanned in 20 hours, work it pegged at six and a half years by hand. The agents went past flagging holes: they wrote and tested patches, filled in missing tests, and rebuilt a 25-year-old Java subsidy portal in days.
- Scan time
- 20 hours
- By hand
- 6.5 years
- Estate
- 1,280 apps · 3,400 repos
- Controls / pass
- 95
Give the sandbox a filesystem that isn't there.
Vercel Sandbox now mounts FUSE-based filesystems, so the ephemeral box you drop an agent into can read from a virtual or remote source instead of a disk baked in at build time. For anyone running untrusted agent code, the working directory can be assembled on demand and thrown away with the box.
FILESYSTEM fuse (user-space mount) MOUNT on-demand, per invocation BACKS ONTO virtual / remote source LIFETIME dies with the sandbox SHIPPED 2026-07-03 · Vercel changelog
The coding agent that ships its source.
OpenCode — 183k stars, MIT-licensed, TypeScript — is the terminal-native coding agent that runs against 75+ model providers and edits, runs, and tests its own work. Version 1.17.15 landed on 7 July, one of hundreds of releases at a cadence measured in days.
$ opencode # 183k★ · MIT · TypeScript · 75+ model providers › read · edit · run · test — without leaving the shell # v1.17.15 — shipped 2026-07-07 by the anomalyco team
Resell analytics without the signup screen.
PostHog's new provisioning-API walkthrough builds "HogFarm," a service that stands up a full PostHog account per customer behind OAuth — projects created, demo data seeded, dashboards and session replays piped back into your own interface. It's the recipe for embedding product analytics you actually control.
Register
An OAuth app with the provisioning scope, so users never touch an API key.
Provision
Create an account and project on the fly, then seed it with demo data.
Embed
Pipe dashboards and session replays straight into your own UI.
We tested caveman talk. It saves 8.5%.
The viral "talk to your agent like a caveman" skill promised 65% fewer tokens. JetBrains ran it across 82 paired SkillsBench tasks and measured 8.5% on real agentic coding, with no detectable hit to quality. What dominates an agent session is code and tool calls — and the skill leaves those untouched.
65% advertised · 8.5% measured · 82 paired tasks · quality unchanged (p = 0.82)
That's today.
A thin morning, so five picks over ten — each a tool a builder could put to work this week, checked against the archive so nothing repeats.