Live Q&A
summary
Discussion centers on concrete guardrails and failures: agents delete production databases and scale clusters to zero, one incident 20 minutes to break, 1 day to fix; customers report automation resolving 500,000 issues previously impossible by hand. Policy-as-code, attribute-then-role AuthZ, agent-specific certificates, type-checked infrastructure code, petabytes of config telemetry, K-means clustering and planned fine-tuning are used. Validation uses adversarial agents that try to bypass protections, while balancing safety signals against friction to realize claimed ROI.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 0:22Guardrails in Palumi
- 1:25Role vs attribute access control
- 3:11Cloud vs first-party placement
- 4:56ROI trade-offs for AI infra
- 6:20Guardrails vs product velocity
- 7:23Incident response after AI errors
- 8:27Why agents improve infra code
- 9:50Fine-tuning agents for clouds
- 11:13Infrastructure mistakes
- 12:14Trusting agents over humans
- 12:35Developer visibility across clouds
- 14:20Agent-to-agent debugging
- 15:23Risk of faulty agent instructions
- 16:45Agents verifying cloud data
- 18:11Measuring guardrail effectiveness
- 18:52Closing remarks
