Zero Trust Is the Foundation AI Agents Need
Anthropic's framework connects agent identity, least agency, tool boundaries, memory integrity, and recovery into one practical security model.
notebook / tag
#agents
6 entries with this tag.
When Agent Autonomy Produces Activity Instead of Work
An audit of Aria's autonomous loop found hundreds of goals, almost no progress, and a completion system that rewarded plausible output.
Your Research Agent Needs a Consumer
Aria produced hundreds of useful research reports that nothing ever read. The fix was a small, idempotent consumer pipeline.
The Observer Changes the Agent
A self-observation feature for Aria showed why metrics should remain available without becoming permanent instructions.
The Fourth-Agent Test
A quick way to tell whether your agent setup is ready to grow or still held together by one-off fixes.
A Small Pattern Library for Tool-Calling Agents
Four simple work patterns for agents that use tools, make changes, test the result, and recover from interruptions.