Internal Background Agent Systems: Architecture and Rollout
summary
Open Inspect runs a Cloudflare-based control plane and data plane split with sandboxes on Modal and Daytona, aiming for reproducible dev environments via warm start snapshots rebuilt ~every 30 minutes so boot plus dependency pull hits ~2–3s (goal <10s). Primary wins are SRE first-response triage, customer-support self-resolution, PMs submitting PRs, and self-serve analytics; costs are tight access control, custom slimmed API tools, and a surprise review bottleneck as PR volume rises—teams use AI code reviewers and Semgrep or risk-based auto-approvals.
AI-written summary · about
chapters (6)
- 0:00Why internal background agent systems are emerging
- 1:14Local machines stop scaling
- 6:08PMs, analysts, and prototypes as first use cases
- 11:44Open Inspect's control plane and sandboxes
- 18:08Deployment lessons after the first rollout
- 22:43Driving adoption and handling review bottlenecks
related talks
26mAgents for the Inner Loop and Outer Loop
36mHow to Safely Deploy AI Agents that Write and Manage your IaC
41mWhat Happens After the POC: Building Agentic Systems That Last
20mThe Missing Layer After Launch
01h 16mAutomating Large Scale Refactors with Parallel Agents
40mWhen Incidents Fix Themselves: AI SRE in action
34mAutomating SRE Operations with Multi-Agent AI
24mAI Governance for developers: A practical framework for deploying autonomous coding agents safely…