AWS, Uber, Anthropic and Google all place agent safety in runtime layers: lifecycle controls, limits on what an agent can read and execute, network isolation at the environment level, and observable behavior. Machine-level secret storage and payment budgets set outside the agent follow the same approach. The practical consequence is that permissions, credentials and spend limits belong in the infrastructure an agent runs in. Prompts and refusal training alone are not enough.
From the Field Notes:
- Trust in Depth: Giving agents more autonomy without losing control (2026-10-21)
- Uber's CISO on building an insider threat program for AI (2026-10-16)
- In the blog post, Anthropic says it identified most of the cases through a review of transcripts that began in July, and says that these cases have minimal real-world impact — less severe from an a... (2026-10-15)
- Developing Gemini 4 Argon responsibly and safely (2026-10-17)
- No more data entry - how to set up your agents to use your passwords. (2026-10-15)
- Building agents that transact with Anchor Browser and AgentCore payments (2026-10-17)