From Experiment To Enterprise: Scaling an AI Agent for Code Review
summary
An internal agent called Code Fusion runs as ephemeral Kubernetes pods to perform bounded reviews across 400+ repositories, producing tens of thousands of code reviews and exporting durable evidence before teardown. An adapter seam routes Anthropic/Cloud Code or OpenAI/Codex engines, scoped identities and admission gates dedupe and gate work, and deterministic scripts handle post-processing to cut inference costs. Evals, release identities, and dual-capacity (K8s vs inference) monitoring detect regressions and keep rollout safe.
AI-written summary · about
chapters (17)
- 0:00Scaling AI agent for code review
- 1:24Parallel AI experiments limits
- 3:30Agent core and enterprise bounds
- 6:21Build internal agentic system
- 7:46Agent workload contract
- 9:53Reuse third-party tooling adapter
- 11:20Enterprise governance security
- 13:07Threat modeling agent security
- 14:34Ephemeral pods enforce workload
- 15:41Detect regressions behavior
- 16:45Preserve evidence for incidents
- 18:12Evaluate agent output quality
- 18:54Uneven usage training churn
- 19:35Admission gates control entry
- 21:43Split judgment and execution
- 23:11Key takeaways scaling AI
- 24:37Q&A on admission gates
related talks
19mNo Token Budget: Agentic Software Factories on Kubernetes
32mFrom Toil To Triumph: Harnessing Agentic AI To Streamline Infrastructure as Code
36mHow to Safely Deploy AI Agents that Write and Manage your IaC
35mThe Agentic AI Engineer
16mWhat Comes After AI Coding Assistants
26mAccelerating DevOps with AI
35mBeyond the Hype and Hyperbole
37mAI-Driven Developer Productivity at Scale