No Token Budget: Agentic Software Factories on Kubernetes
summary
OpenAI runs an exploratory system where agents own the full lifecycle: agents write, review, and release code and run 24/7 for months without human intervention, including rollbacks and production remediation. Architecture uses an actor pattern, a control plane and execution plane that materialize heavyweight runtimes, and hermetic development environments (OCI containers, isolated k8s, Playwright) so candidates are full deployments. Optimization uses SPOs where SPO graders compare deployed instances and only accept candidates that measurably beat master (for example better-than-1ms latency). Lessons: models are capable; infrastructure and hermeticism decide success.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:15Project roadmap
- 1:52The greenfield premise
- 3:28Six constraints for an autonomous software factory
- 5:16Modeling the lifecycle with actors
- 6:08Tell, spawn, and listen
- 6:55Control and execution planes
- 8:50Hermetic development environments
- 9:50Single-problem optimization
- 11:16Infinite and finite alignment
- 13:16Agent-run CI/CD and SRE
- 14:40Get out of the models' way
- 15:50Why hermeticism wins
- 16:50Reducing the solution space
- 17:24Managing code entropy continuously
- 18:00Closing lesson
related talks
52mDesigning Production-Ready Agentic Systems: Lessons from the Field
17mAI TechOps: Data and Infra for AI Foundation to Scale and Manage 1000s of Cohesive Agents
27mFrom Experiment To Enterprise: Scaling an AI Agent for Code Review
26mBeyond Sandboxes: Architecting Durable Runtimes for AI Agents
10mKeynote: The Open Source Runway: Building the Foundation for the Agentic Era
36mHow to Safely Deploy AI Agents that Write and Manage your IaC
31mBuilding and Orchestrating Real Production-ready Agentic AI Systems
27mFrom Chatbots to Agents: Orchestrating Agentic AI on Existing Enterprise Kubernetes Infrastructure.