Session

Secure, Portable, Collaborative: Multi-Harness Agent Teams with Omnigent

Overview

Experience	In Person
Track	Artificial Intelligence & Agents
Industry	Enterprise Technology
Technologies	Databricks Agents
Skill Level	Advanced

Agent work is moving beyond a single developer prompting a single model on a laptop. Omnigent provides the capabilities needed for team-scale agent work: portable context across harnesses, reusable agents and tools, session history, policy and quality protocols, async approvals, and mobile and web UI. This deep dive shows how those pieces fit together to make agent work persistent, collaborative, and production-ready. We'll show this in both forms: Omnigent OSS, the open meta-harness you can run anywhere, and Omnigent on Databricks, where the same agents, skills, and policies run on top of the governed Databricks Platform.

Session Speakers

Elise Gonzales

/Staff Product Manager
Databricks

Kasey Uhlenhuth

/Director, Product Management
Databricks

Full Summary

Omnigent: A meta harness for composing, controlling, and collaborating with coding agents

Coding agents are now common in daily development, yet they often run in silos with brittle autonomy controls and awkward handoffs between tools and teammates. Omnigent, an open source project from Databricks, introduces a meta harness layer that sits above individual coding harnesses such as Claude Code and Codex, unifying orchestration, guardrails, and shared context. The goal aligns with emerging agentic development patterns on the Databricks Platform: multiple specialized agents working together, governed, and observable. Teams exploring how this fits into a broader strategy will find grounding in what it means to build a data and AI strategy around platform focus.

The session showcases the open source experience, then the managed version integrated with Databricks sandboxes and Unity AI Gateway. The result is a single place to run, review, and refine multi agent coding work, with collaboration built in and model governance applied automatically.

Why an orchestration layer above coding agents is needed

Developers struggled with three recurring pain points. Composition was hard, with parallel tmux or custom dashboards to track separate sessions across harnesses, followed by duplicated work as new harnesses were adopted. Autonomy was blunt, swinging between prompting for every action or granting sweeping permission, which made engineers either micromanage or over trust. Collaboration was clumsy, with copy paste between agents, teammates, and plan files, and no simple way to view or discuss what an agent changed.

Omnigent addresses those gaps by acting as a harness of harnesses. It wraps existing coding agents and exposes them through a native app and web UI with shared context, markdown viewing, an integrated shell, and the option to switch harnesses mid session. Engineers get a coherent workspace where multiple agents can be composed without losing state or visibility.

Multi agent skills, cross review, and shared sessions

Two built in skills illustrate what a meta harness unlocks. Polly is a multi agent coder that supports cross review, fan out for parallel tasks, and read only investigation for audits and debugging. Having one harness implement and another review consistently yielded higher quality than self review by a single agent. Debbie, a multi agent debate skill, stress tests ideas without writing code.

Collaboration is treated as a first class capability. Sessions can be shared with read or edit permissions, so a teammate can join an active run, see recent activity and working files, and add context or changes before a pull request exists. Pair programming and code reviews move into the shared session, reducing friction and eliminating much of the back and forth that previously lived in terminals and DMs.

Managed Omnigent: cloud runners, governance, and intelligent routing

The managed version adds two capabilities that are difficult to replicate locally. First, the runner can live in a Databricks sandbox that spins up in roughly a second, so agents keep working after the laptop closes and can be monitored from a phone. Sandboxes host agents securely, are easy to create and destroy, and let developers take sessions on the go.

Second, Omnigent connects to Unity AI Gateway for governed model access, budgets, and policy enforcement. Budgets become live inputs to routing and usage decisions, with policies that can warn, guide, or shift behavior as spend approaches limits. A ucode style budget view shows real time spend and policy thresholds, encouraging intentional model choice without blocking work.

Those pieces enable intelligent routing at the meta harness layer. Omnigent inspects sub tasks and routes to different model tiers automatically, using fast open source models like Kimi for cheap mechanical lookups, mid tier options such as Gemini Flash for edge case review, and top tier models for deep audits. Intelligent routing depends on the budget and capacity signals in Unity AI Gateway, so it is currently a managed Omnigent capability.

Open source first, built in the open

Omnigent was built with Omnigent, and the team open sourced it to keep the orchestration layer above coding agents vendor neutral. Following the path of Spark and MLflow, development is happening in public with an active Discord community and external contributors already landing improvements. The open source project is available today, and managed Omnigent is in beta for Databricks workspaces.

From isolated agents to governed, collaborative swarms

Real leverage with coding agents comes from composition, control, and collaboration, not picking a single favorite model. Omnigent turns fragmented agent work into shared, auditable workflows, and the managed version adds cloud persistence and policy aware routing so teams can scale quality without losing guardrails.

FAQ

No. Omnigent is open source and runs locally with your own harnesses and models. The managed version adds Databricks sandboxes, Unity AI Gateway integration, and intelligent routing that uses budget and policy signals.

Having one harness implement and a different harness review introduces independent perspective and reduces blind spots. The team consistently observed higher quality than self review by a single agent, at the cost of invoking more than one provider.

The meta harness inspects the sub task and routes to an appropriate model tier, for example Kimi for quick lookups, a mid tier model for edge case review, and a top tier model for deep audits. Budget state and policies from Unity AI Gateway can shift routing as spend approaches thresholds.

With managed Omnigent running in a Databricks sandbox, the runner continues in the cloud. Agents keep executing and you can rejoin from another device, including a phone.

You can share a session with read or edit permissions. Collaborators see the active agent, recent conversation, and working files, and can add context or make changes before any pull request is opened.