Omnigent Adds Contextual Policies for Enhanced AI Agent Governance
Omnigent, an open-source meta-harness for AI agents, now features contextual policies that leverage session state to enhance governance. These policies allow for more nuanced control over agent actions, improving security and cost management by considering the agent's history within a session. This capability is available for various coding and custom agents wrapped by Omnigent, providing richer policy options than traditional static controls.
- →Contextual policies for AI agents track session state
- →Enhanced governance controls for agent security and cost
- →Session state enables dynamic policy enforcement
- →Google Drive policy scopes access based on session activity
- →Risk score policy dynamically adjusts agent behavior
Features (6) ›
- Contextual policies for AI agents track session state
Omnigent introduces contextual policies that can track session history, such as documents accessed or costs incurred, to make dynamic decisions about subsequent actions. This enables more sophisticated governance than static allow-lists or deny-lists.
- Enhanced governance controls for agent security and cost
Contextual policies allow for granular control, such as per-session spending caps or guardrails that adapt to accumulated risk. This addresses new enterprise risks introduced by AI agents, like prompt injection and sensitive data access.
- Session state enables dynamic policy enforcement
Policies can update arbitrary session-specific variables based on agent events like tool calls or LLM interactions, allowing for context-aware decisions on allowing, denying, or transforming agent actions.
- Google Drive policy scopes access based on session activity
An example policy confines agent writes to documents created within the current session and can tighten restrictions if the agent accesses confidential material.
- Risk score policy dynamically adjusts agent behavior
This policy maintains a running risk score for a session, prompting user approval for sensitive actions once a configured threshold is crossed.
- Budget policy enforces spending caps on agent sessions
A budget policy tracks model call expenditures and can pause or block further actions once spending thresholds are met.
https://www.databricks.com/blog/contextual-policies-omnigent-using-session-state-better-govern-ai-agents