Databricks Open-Sources Omnigent Meta-Harness for AI Agents

Databricks has announced the open-sourcing of Omnigent, a new "meta-harness" for AI agents. Released under the Apache 2.0 license, Omnigent is currently in its alpha stage and was developed by the Databricks AI team with Neon.

Omnigent functions as an interface that sits above existing coding agents such as Claude Code, Codex, and Pi, as well as SDKs like OpenAI Agents and the Claude Agents SDK. Its core design principle is to standardize the user-facing interface of various harnesses, making them interchangeable components within a larger system. This allows engineers to manage multiple agents from a single point, addressing the common practice of juggling several agents and manually transferring information between them.

Key capabilities of Omnigent include composition, control, and collaboration. Composition enables users to combine different models, harnesses, and techniques without rewriting code, facilitating easy switching between agents. Control is provided through stateful, contextual policies that enforce guardrails at the meta-harness layer, such as pausing an agent after a certain cost threshold or requiring human approval for critical actions like a `git push` after package installation. Collaboration features allow for sharing live agent sessions via URL, enabling teammates to observe, chat, comment, co-drive, or fork conversations in real-time. An underlying OS sandbox, Omnibox, enhances security by locking down OS access and managing network requests.

The architecture of Omnigent consists of two main parts: a runner, which wraps any agent in a sandboxed session with a uniform API, and a server, which manages policies and sharing. The server exposes every session across terminal, application, and web APIs, ensuring messages, sub-agents, terminals, and files remain synchronized across devices.

Examples of agents shipping with the repository include Polly, a multi-agent coding orchestrator that plans and delegates work to sub-agents in parallel git worktrees, and Debby, a brainstorming partner that leverages both Claude and GPT to critique answers. These examples illustrate how Omnigent can coordinate different large language models for tasks like planning, search, and code generation.

According to Marktechpost, an interactive demo has been created to showcase Omnigent's meta-harness workflow, featuring the Polly orchestrator delegating tasks to Claude Code, Codex, and Pi simultaneously.