Introducing Dogwood: runtime verification for AI agents

This article introduces Dogwood, an open-source governance language designed to enforce safety policies on AI agent tool calls by examining sequences of actions rather than isolated requests.

• Extends Cedar policy language with temporal conditions that reference event history, not just current requests
• Supports policies requiring approval before actions, rate limiting, and preventing tool use after accessing sensitive data
• Provides operators like "formerly," "count_within," "count_distinct_within," and "sum_within" for common policy patterns
• Fully backward compatible with existing Cedar policies; no migration required
• Built on Metric First-Order Temporal Logic (MFOTL) from runtime verification discipline
• Available open-source under Apache 2.0 license with language guide and reference implementation

Dogwood enables teams to govern complex agent workflows by expressing rules about action sequences, prerequisites, and resource limits while maintaining Cedar's authorization guarantees.