Omnigent, an open source meta-harness for AI agents from Databricks, introduces contextual policies that track session state to make governance decisions. Unlike simple allow/deny rules, these policies remember what an agent has done within a session — which documents it read, how much it spent, what the user originally asked — and use that accumulated context to decide whether the next action should proceed. Three built-in policies are included: a Google Drive policy that restricts writes to documents the agent created in the current session and implements Bell-LaPadula no-write-down rules for confidential files; a risk-scoring policy that raises a running score as sensitive actions accumulate and triggers human approval once a threshold is crossed; and a budget policy that pauses or blocks expensive model calls when per-session or per-user spending caps are hit. A fourth example, intent-based authorization, locks each session to only the tools relevant to the user's original prompt, implementing least-privilege automatically. Omnigent wraps existing agent harnesses including Claude Code, Codex, and custom OpenAI/Claude SDK agents, applying these policies uniformly across all of them.
