Databricks just open-sourced the layer above every agent harness

Databricks open-sourced Omnigent today. A meta-harness that composes Claude Code, Codex, Pi, and custom agents into one unified interface.

This is not another agent framework.

The insight is structural: every agent harness wraps an LLM into a silo. Claude Code talks to Claude. Codex talks to GPT. Pi talks to Pi. None of them talk to each other. Engineers end up with 4-5 agents open, copy-pasting between them. Omnigent sits above all of them.

How it works: a runner wraps any agent in a sandboxed session with a uniform API. A server provides policies and sharing. One command starts a session in terminal. That same session is live in browser. On your phone. Shared with a teammate.

The architecture decisions matter:

- Composition without rewriting. Switch between Claude Code, Codex, Pi with one-line YAML changes. Mix harnesses in the same agent.
- Contextual policies, not prompts. Track dynamic state per session. After an agent downloads an npm package, require human approval to git push. After $100 in LLM spend, pause and ask.
- Real-time collaboration. Share a session via URL. Teammates watch your agent work. Co-drive on your machine. Fork conversations.
- OS sandbox with network interception. The agent never sees your GitHub token. The egress proxy injects it only on approved requests.

This is the Kubernetes moment for agents. Each harness is its own silo — its own context, controls, running model. None of it carries over when you switch tools. Meta-harness lifts your work above any single harness so sessions, policies, and skills stay with you regardless of which agent or model is running.

106 stars in hours. Apache 2.0. GitHub: github.com/omnigent-ai/omnigent

Matei Zaharia (Apache Spark co-creator) is building this at Databricks with a 5000-person engineering team as the test bed.

If you are running multiple agent harnesses today, audit your stack. The meta-harness layer is arriving whether you build it or adopt it.