Databricks Open-Sources Omnigent Meta-Harness for AI Agent Composition

Databricks has open-sourced Omnigent, an Apache 2.0 "meta-harness" for composing and governing AI agents. The tool provides a unified interface for agents like Claude Code, Codex, and Pi, making them swappable. Key features include stateful control policies for cost and security, and real-time collaboration via shared sessions. Example agents like Polly, a coding orchestrator, demonstrate its capabilities.

Omnigent meta-harness release

• ▪Roadmap items for Omnigent, such as the Omnigent Server MCP, are not yet shipped
• ▪The Omnigent project is released under the Apache 2.0 license and is currently in an alpha stage
• ▪Databricks has released Omnigent, an open-source "meta-harness" for AI agents

Architecture with runner server

• ▪Omnigent's architecture consists of a "runner" that wraps agents in a sandboxed session and a "server" that provides policies and sharing
• ▪An OS sandbox called Omnibox is used to lock down OS access and can keep secrets like GitHub tokens hidden from the agent
• ▪The Omnigent server exposes each session through the terminal, an app, and web APIs, including a local web UI at localhost:6767

Composition across multiple harnesses

• ▪The "composition" feature allows combining different models and harnesses, and switching between agents with one-line code changes
• ▪Omnigent provides a common interface for agents like Claude Code, Codex, and Pi, making them swappable components

Control through stateful policies

• ▪Example policies can pause an agent after it exceeds a cost budget or require human approval for a 'git push' command after an npm install
• ▪Omnigent's "control" feature uses stateful, contextual policies to enforce guardrails at the meta-harness layer, not through prompts

Collaboration via shared sessions

• ▪In a shared session, teammates can co-drive the agent, comment on files, or fork the conversation
• ▪The "collaboration" feature allows users to share live agent sessions via a URL for real-time viewing and interaction

Example agents Polly Debby

• ▪Another example agent, Debby, is a brainstorming tool that uses both Claude and GPT models, showing their answers side-by-side
• ▪The Polly agent plans tasks, delegates work to sub-agents, and routes code for review by an agent from a different vendor
• ▪Omnigent ships with an example agent named Polly, which acts as a multi-agent coding orchestrator