Anatomy of a 1,000-agent run: sandboxing, blocking classifiers, and offline monitoring
RLanceMartin · x · 2026-11-10
Lance Martin outlines the safety approach of the 1,000-agent experiment: sandboxing, real-time monitoring, and offline monitoring.
The architecture diagram adds specifics: each sandbox is its own lightweight VM with no network or credentials, reachable only via git and SQL pipes; a gateway runs a fail-closed blocking classifier on every tool call; the git repo accepts own-branch pushes only, and the shared Postgres role has time and size limits; run records live outside the sandboxes for offline review, with serious findings escalated to run owners.
More from coding & agent
- The One-Person Million-Dollar AI Agency: An Agent-Stack Blueprint With the Math — boringmarketer · 2026-11-10
- Developer natively manages 25-30 parallel agent sessions on macOS with Rex — jonathan_wilke · 2026-11-10
- Pi Dash orchestrates 10 Muse Code agents with Muse Ember, rewriting 178K lines in a day — AIatMeta · 2026-11-10
- Fable 5.2 builds a voxel destruction engine from a single prompt, runs on a MacBook — mattshumer_ · 2026-11-10
- Microsoft Foundry extension for azd hits GA, agent config lives in azure.yaml — WirelessLife · 2026-11-10
- Microsoft's BoxBox lets agents hot-reload their own harness mid-run, hitting 1.70x speedup — dair_ai · 2026-11-10