Databricks Open-Sources Omnigent, A Meta-Harness For Managing Multiple AI Agents

Databricks has released Omnigent, an open-source framework designed to sit above existing AI agent tools and give organisations a single layer for orchestrating, governing and collaborating across multiple agents.

The project, released under an Apache 2.0 licence in alpha, is intended to address what Databricks views as a growing friction point: as organisations adopt more AI agent tools – from coding assistants like Claude Code and Codex to custom-built internal agents – those tools operate in silos with incompatible interfaces, making it difficult to combine them, swap between them or apply consistent controls.

Omnigent introduces what Databricks calls a “meta-harness,” a layer that wraps around individual agent tools and exposes them through a common API. The underlying logic is that regardless of how each agent harness communicates with its language model internally, the user-facing interface follows the same pattern: messages and files in, text streams and tool calls out.

The project was authored by Matei Zaharia, Kasey Uhlenhuth and Corey Zumar from Databricks’ AI team. Zaharia is Co-Founder and CTO of Databricks.

The problem Omnigent targets

Databricks adopted coding agents across its 5,000-strong engineering organisation and has built agents for customers at scale. That experience, the company argues, has shown that agent engineering is moving beyond what a single harness can handle on its own.

The team pointed to several examples of multi-agent patterns already in production use. Legal AI firm Harvey pairs an open-source worker model with a frontier advisor model to manage quality and cost. Anthropic built its research product using a lead agent that orchestrates parallel sub-agents. Databricks’ own Genie product uses different language models for planning, search and code generation.

These patterns span multiple harnesses, models and teams, but each harness only understands its own sessions. Combining agents, governing them and collaborating on them with colleagues requires something above the individual harness, which is the gap Omnigent is designed to fill.

The Omnigent team described the current state as comparable to an earlier era of infrastructure management, before orchestration tools like Kubernetes and Terraform allowed engineers to manage fleets of servers rather than individual processes.

“Each harness is its own silo, with its own context, its own controls, and its own way of running, and none of it carries over when you switch tools,” the authors wrote.

How Omnigent works

Omnigent’s architecture centres on a “runner” component that wraps any agent in a sandboxed session with a uniform API. A server layer sits above the runner to handle policies and sharing, and exposes each session through a terminal interface, a desktop or web application and web APIs.

The framework supports both terminal-based coding agents such as Claude Code, Codex and Pi, and agent SDKs including OpenAI Agents and the Claude Agents SDK. Switching between agents requires a one-line change in a YAML configuration file, and users can combine sub-agents from different harnesses within a single workflow.

Agents can run locally or be launched on hosted sandbox providers such as Modal and Daytona for isolated, collaborative environments.

Governance and cost controls

One of Omnigent’s central features is its approach to security and cost policies, which operate at the meta-harness layer rather than being enforced through prompts.

The framework supports contextual security policies that track dynamic state within each session. For example, an organisation could configure a rule requiring human approval for a git push after an agent downloads a new npm package, or restrict an agent to writing only to documents it created.

Cost controls allow organisations to set spending thresholds per session. Omnigent tracks LLM costs dynamically and can pause an agent and prompt for approval after reaching a defined limit, such as $100 per session.

The system also includes an OS-level sandbox built by Databricks’ security team, with the ability to lock down operating system access and intercept network requests. One example given is preventing an agent from ever seeing a GitHub security token directly, instead injecting the token via an egress proxy only on approved requests.

Real-time collaboration

Omnigent allows users to share live agent sessions via URL. Team members who join a shared session can observe what an agent is doing, comment on files in its workspace or send commands directly.

Once an agent is connected to the Omnigent server, it becomes accessible through multiple interfaces – web, mobile, a native macOS application and APIs – meaning team members can follow a session across devices.

Databricks frames this as a shift in how teams work with agents, with the agent’s session and working directory becoming the central place for collaboration rather than a separate tool like a shared document or messaging channel.

Roadmap

Databricks has published a roadmap for Omnigent that includes automatic optimisation at the meta-harness level using a technique called GEPA, code-based introspection within agents drawing on approaches like MemEx and RLM, and an Omnigent Server MCP so agents can operate across sessions.

The framework currently supports deployment on Fly.io, Railway, Modal and Daytona sandboxes along with a range of LLM providers, and the team is inviting contributions for further integrations.

Omnigent is available on GitHub with a quickstart guide and documentation. The project also has a Discord community for contributors and users.

Last Updated on June 15, 2026 by Nick Ross