Databricks has launched Omnigent, an open source tool the company describes as a meta-harness for managing AI agents across different systems. The project is intended to help enterprises combine, control, and share agent workflows without being locked into a single agent framework.
The company said Omnigent sits above existing agents such as Claude Code, Codex, Pi, and custom-built tools, giving users a common way to work with them. Databricks is releasing the software under the Apache 2.0 license.
The move reflects a broader shift in how companies are using AI agents. Databricks said that in its own engineering organization, which has more than 5,000 employees, teams now work with multiple agents at once and increasingly design workflows that coordinate several models and tools. The company argues that these patterns create problems that a single agent harness cannot solve on its own.
Databricks says Omnigent is meant to address three main challenges: composition, control, and collaboration.
On the composition side, the platform is built to let teams combine different models, harnesses, and techniques without rewriting code. The company says users can switch between agent systems with a one-line change, which could make it easier to adapt as tools evolve.
For control, Omnigent adds policy features that operate at the meta-harness level rather than relying only on prompts. Databricks says these policies can track session state and enforce rules such as cost limits or permission checks. For example, an organization could require human approval before an agent pushes code after downloading a package, or limit an agent to editing only the documents it created.
The collaboration features are aimed at making agent sessions more shareable. Databricks says users can invite teammates into a live session, let them comment on files in the workspace, and even send commands while the agent is running. Sessions can also be accessed through links, giving teams a shared workspace around the agent itself.
Omnigent wraps terminal-based coding agents and agent SDKs in a common interface, according to Databricks. The company says the system can expose the same agent session through the web, a mobile experience, a native Mac app, or APIs.
It also supports running agents locally or in hosted sandboxes from third-party providers. Databricks said that approach is meant to make collaboration safer by keeping sessions in a controlled environment.
Security is another central part of the pitch. Databricks says Omnigent includes a strong operating system sandbox and can intercept or alter network requests to reduce risk. The company gave the example of preventing an agent from directly seeing a GitHub security token and only providing it through a proxy for approved actions.
The platform also includes cost controls that monitor spend dynamically. Databricks said a user could configure the system to pause an agent after it reaches a certain spending threshold and ask for approval to continue.
Databricks is positioning Omnigent as an early step toward a new layer of software abstraction for agent workflows, similar to how cloud orchestration tools changed the way engineers managed infrastructure.
The company said Omnigent is in alpha and is now open source. It also said it plans to add more capabilities over time, including better optimization features, deeper agent introspection, and support for additional harnesses and integrations.
For now, Databricks is encouraging developers to test the project, review the documentation, and contribute to the codebase as it evolves.