Databricks has introduced contextual policies in Omnigent, giving teams a more flexible way to govern AI agents by using session history, not just one-off allow or deny rules, to decide what an agent should do next.
Whatโs new
- Policies that remember the session: Omnigent policies can track what an agent has done so far and use that context to make smarter decisions about the next action.
- Works across existing agent harnesses: Because Omnigent is a meta-harness, the same policy layer can be applied across coding agents like Claude Code, Codex, Goose, and Hermes, as well as supported custom agents.
- Built-in cost controls: Omnigent can track session-level LLM cost and pause an agent to ask whether it should continue after crossing a spend threshold.
- Stronger guardrails as risk builds: Instead of treating every action in isolation, Omnigent is designed to support policies that get stricter as more risk accumulates in a session.
- Custom Python policies are supported: Teams can write their own policy functions in Python, register them on the server, and use them alongside the built-in policies.
Instead of only checking one action at a time, Omnigent can evaluate actions in the context of the full session, which makes policies like per-session budgets, approval thresholds, and more context-aware security controls possible.
๐ Read the full post here
