Databricks Open Sources Omnigent to Put a "Meta-Harness" Above AI Agents

Why it matters

OpenSharing is Databricks' attempt to set the protocol layer for enterprise AI asset exchange before proprietary agent marketplaces harden around the market.

Matei Zaharia (@matei_zaharia), Databricks' co-founder and CTO, and Kasey Uhlenhuth are open sourcing Omnigent, an Apache 2.0 alpha project that Databricks describes as a "meta-harness" for AI agents.

The pitch is not another model wrapper. Omnigent is meant to sit above the agent harnesses engineers already use, including Claude Code, Codex, the Pi agent and custom agents, and make them composable, controllable and shareable through one interface. Zaharia said in an X thread that Databricks saw engineers at Databricks and Neon combining multiple agents into loops and workflows, but struggling because each harness kept its own interface, session state and controls.

That is the story: Databricks is trying to define the layer above the coding-agent boom before the workflow hardens into a set of incompatible tools.

Omnigent targets three problems that sit above any single harness: composition, collaboration and control. For composition, the project provides a uniform API over command-line agents and agent SDKs, so developers can build multi-agent teams that mix harnesses and models, or swap a harness and model mid-session or mid-loop. Databricks frames that as the next abstraction after agent harnesses made models swappable.

For collaboration, Omnigent adds common terminal, web, desktop and mobile interfaces over the same agent session. The Databricks blog says users can invite teammates to view a live session, comment on files in the workspace or send commands. Zaharia's thread frames the problem more bluntly: engineers are copy-pasting text and screenshots between agents, Slack and Docs because their agent sessions are not themselves collaborative workspaces.

For control, Omnigent introduces contextual policies that track session state rather than relying only on static allow-or-deny prompts. The examples in the Databricks post are specific to the risks of agentic coding: require approval before a git push if an agent downloaded an untrusted npm package, restrict an agent to docs it created, or pause and ask after every $100 of LLM spend. Databricks also says Omnigent includes a strong OS sandbox from its security team, with controls for OS access and network requests, including an egress proxy pattern that can keep a GitHub security token out of the agent's direct context.

The project reflects Databricks' own internal pressure. The company says it adopted coding agents early across its 5000+ member engineering team and has built thousands of agents for customers. In the blog post, Zaharia and Uhlenhuth argue that the best agent results increasingly come from systems rather than a single model in a single harness, citing patterns such as a lead agent orchestrating parallel subagents and Databricks' Genie using different LLMs for planning, search and code generation.

Omnigent's core abstraction is that, whatever happens inside each harness, the user-level interface has a common shape: messages and files go in, text streams and tool calls come out. Databricks is wrapping that shape in a runner, server and shared interfaces. The architecture described in the post puts agents inside sandboxed sessions with a uniform API, while the server provides policies and sharing and exposes sessions across terminal, app and web APIs.

The deployment story is deliberately broad. Zaharia said Omnigent can run on a laptop or be deployed to Docker, Railway and Fly.io, and can run agents in cloud sandboxes on Modal and Daytona. The project works with LLM providers and coding-agent subscriptions, including OpenRouter and Databricks. Databricks is also pointing developers to Omnigent's quickstart, documentation and GitHub repository.

The strategic move is familiar for Zaharia. Apache Spark helped Databricks turn an open source infrastructure primitive into a commercial platform company. Omnigent is an attempt to do something similar one layer above agent harnesses: make the shared control plane open, then let the commercial and ecosystem gravity gather around the workflows it standardizes.

There are limits baked into the launch. Omnigent is alpha today, and Databricks describes several important ideas as roadmap items, including meta-optimization with GEPA, programmatic tool calling similar to RLM and MemEx, an Omnigent Server MCP so agents can work across sessions, and more harness integrations. The claim that the field needs a meta-harness is plausible, but the implementation still has to prove that developers will route their daily agent work through a new layer rather than stay inside the native tools from Anthropic, OpenAI and others.

That tension is what makes Omnigent worth watching. AI coding agents are becoming both more powerful and more fragmented. Databricks is betting that the next durable primitive is not another agent, but the layer that lets teams combine agents, enforce policies across them and work together inside live sessions. If that layer becomes real, the agent market starts to look less like a collection of chat windows and more like infrastructure.