Skip to content

SYSTEM Cited by 4 sources

Omnigent

Omnigent is Databricks' open-source AI agent framework, notable for its built-in contextual policy engine that evaluates every tool call against multiple policy layers before allowing execution.

Architecture

The framework runs a single policy engine per session that evaluates tool calls against a stack of contextual policies. Each policy returns one of three verdicts:

  • ALLOW — action proceeds.
  • ASK — action is paused for human approval.
  • DENY — action is blocked.

Policies compose with deny-wins semantics: if any single policy denies an action, it is blocked regardless of what other policies say. This makes the policy set monotonically restrictive — adding a new permissive policy cannot override an existing denial.

Built-in policy types

  • Intent-based authorization — binds a session to a declared purpose; any tool call outside that purpose is denied or gated (intent-based-authorization).
  • Session-risk scoring — behavioral anomaly detection that tracks drift from expected session patterns (described in companion blog "Blocking Slow-Burn Attacks").
  • PII blocking — prevents exposure of sensitive data in agent outputs.
  • Custom rules — user-defined policies in plain language, compiled to enforceable rules.

Key design decisions

  1. Intent is immutable at runtime. For autonomous agents, intent is pinned in the agent spec. For interactive agents, intent is set at session start with human approval. The running agent has no tool to edit, remove, or disable policies.

  2. Adding policies requires human approval. Even if an agent has a "add policy" tool, a built-in meta-rule gates new policy activation on explicit user consent.

  3. Deny-wins composition prevents policy escalation — a new permissive rule cannot lift an existing block (deny-wins-policy-composition).

Meta-harness role (2026-08-07 disclosure)

Beyond its policy engine, Omnigent is Databricks' meta-harness — the end-user coding tool that surfaces a single UX to developers while dispatching to underlying harnesses (proprietary and open-source). It is the default mode for Databricks developers. In the AI-coding cost-management playbook (sources/2026-08-07-databricks-managing-ai-coding-costs-at-scale), the meta-harness is the client-side lever that preserves model flexibility (so spend can move to newer efficiency-frontier models without forcing a per-developer harness migration) and enables task-level routing — examining a task's complexity and delegating the whole end-to-end task to an appropriately-sized model (patterns/specialized-agent-decomposition).

Omnigent (client-side meta-harness) and Unity AI Gateway (server-side proxy/router) are Databricks' two open-sourced/free cost-management components — the pair covers both the meta-harness and AI Gateway halves of the stack.

Smart Routing inside Omnigent (2026-08-13 disclosure)

The 2026-08-13 Smart Routing post (sources/2026-08-13-databricks-smart-routing-in-unity-ai-gateway) specifies how Omnigent operationalizes routing:

  • Developers can select "Smart Routing" in Omnigent instead of manually choosing a harness. Omnigent then selects both the harness and the model per task, with model routing powered by Unity AI Gateway Smart Routing — the concrete instantiation of patterns/specialized-agent-decomposition with the model-selection half now specified (patterns/two-stage-evaluation).
  • Because it routes at task granularity, Omnigent respects task-aware routing — a model is committed for the session to preserve cache hit rate.
  • All sub-agent launches go through the Smart Routing API. Sub-agents can get a different harness+model with a fresh cache and clear instructions — useful because the user's initial prompt is often underspecified. A single task can span nuanced routing across planning + parallel sub-agents (e.g. cheap models for large-codebase summarization, expensive models for architecture design), compounding savings.
  • Admin/org customizations (org-level guidance, whether to reuse prior conversation history) are applied by Omnigent without changing the client.

Status

Open source, in alpha as of July 2026. Documentation at omnigent.ai/docs.

Seen in

Last updated · 766 distilled / 2,225 read