Challenges of Ambient Agents: Building Governable AI Systems at Scale
summary
Ambient agents must be fast, cheap, and governable simultaneously, so the concrete solution is to convert raw event streams into actionable state using stream tools like Flink, Materialize, Redpanda/Kafka (KSQL), plus sketches, t-digest and DuckDB to compute aggregates, anomalies and Minimal Viable Context before calling an LLM. Identity and scoped delegation rely on sender-constrained tokens (DPoP/mTLS), short‑lived credentials and token vaults, while audit‑mode discovery, sampling and behavioral policy layers (Marshmallow AI, Milpond-style prompt guards) enforce norms via a governor pattern that limits drift and unexpected actions.
AI-written summary · about
chapters (8)
- 0:00Introduction to ambient agents
- 1:57Why ambient agents are challenging
- 4:47Context engineering for agents
- 11:35Governance and security risks
- 16:51Identity and access control
- 20:46Complementary security solutions
- 23:59Scaling governance with audit history
- 26:50Q&A
related talks
37mContext is King in AI Agents
37mBringing stories to life with AI
21mDemystifying AI Agents: A New Primitive for Flow Control
29mBeyond Request/Response: What AI Agents Actually Need From Your Event Streams
41mBringing stories to life with AI, data streaming and generative agents
40mIn AI We Trust? Securing the Future, One Agent at a Time
24mAI Governance for developers: A practical framework for deploying autonomous coding agents safely…
44mBuilding Agentic AI Systems: From Transactions To Conversations