Sharing the world

Block’s answer was dedicated cloud workspaces, one isolated environment per agent. Quoted from YouTube’s auto-generated captions.

Worktrees separate the code, not the state. Cursor runs each cloud agent in its own VM with its own database, and he guesses they run “multiple thousands a day”. Quoted from YouTube’s auto-generated captions.

The old pattern was a few shared test environments, one per tester, with people coordinating who had the system. Agents that check their own work need environments that are created and torn down per change. Quoted from YouTube’s auto-generated captions.

Isolating the code also isolated the coordination. Shown by reading the source, not from an incident where two agents took the same task. The fix, PR #3210, was still unmerged in October 2026.

Giving each worker its own branch let them write without contention, but they couldn't coordinate. The design doc moved everything to one shared main with explicit transactions, live since 24 February 2026.

On one 12-repository workspace this left 109 state directories (97 stray) and about 25 daemons. Fixed in the reporter's fork; the upstream issue is still open. A 2025 issue (#765) describes the same root cause.

The fix was to make the app bootable per worktree, with an ephemeral logs, metrics and traces stack, so Codex could run and check one instance per change itself.

Isolation is not only about files and environments. A shared model quota is a resource agents compete for too.

What a naive retry does after the coordinator dies mid-handoff. She also lists stranded claims, duplicate sessions after retries and stale completions arriving after a newer generation started.

At 200+ agent PRs a day, a full environment per branch was too expensive. Each testbox overrides only the service under change and gets private copies of the datastores it declares.

Isolation that leaks fails silently. They added a capture agent that checks the connections each test actually makes against the service’s declared manifest.

The four are Stripe’s Minions, Ramp’s Inspect, Spotify’s Honk and Paxos’s own Hoplites. A useful checklist of what company agent factories have converged on.

Each minion gets a pre-warmed devbox, ready in about 10 seconds with Stripe code and services loaded and cut off from production and the internet. Over a thousand merged PRs a week are written entirely by minions.

Stripe keeps a warm pool of identical devboxes, with repositories cloned and Bazel and type-checking caches already warm, and gives each task its own box.

The same section adds: "Only the integration owner edits shared manifests and lockfiles or reconciles overlapping changes."

Every Inspect session starts from a repository image rebuilt every 30 minutes, so isolation comes with a known staleness bound. Inspect writes about 30% of merged PRs in Ramp’s frontend and backend repositories. The Pragmatic Engineer later reported 75%.