Skip to main content

Reports

Long-form writing on AI systems work: what breaks after the demo, what to measure, and how to leave the system operable.

3 reports in Multi-agent systems

Latest in this domain

Multi-agent systems7 min read

Task-based access control for tool-calling agents

A user role that can 'use support tools' is not the same as permission to refund on a chat that only asked for status.

  • Derive or declare a task object, then authorize tools against that task.
  • Re-evaluate authorization as the conversation drifts across turns.
  • Deny irrelevant tools even when the user's broad role would allow them.
Read report