Databricks open-sourced Omnigent, an Apache 2.0 project it calls a meta-harness, and the distinction in that word is the whole idea. A harness is what wraps a model and turns it into an agent: Claude Code, OpenAI Codex, and Pi are each harnesses. Omnigent sits one level above them and treats each harness as an interchangeable part of a larger system, standardizing the interface so you can swap between agent systems without rewriting anything. The motivation, in the team's own words, is the mess most people are already in: "Many engineers now juggle four or five agents at once. They copy text between coding agents, search tools, Docs, and Slack. Each harness only understands its own sessions." Omnigent adds the shared layer the harnesses lack. Architecturally it is two pieces, a Runner that wraps any agent in a sandboxed session behind a uniform API, and a Server that holds policies and sharing; one command starts a terminal session and a local web UI at localhost:6767 that then syncs across browser and phone.
The pitch organizes around three C's. Composition: switch between Claude Code, Codex, Pi, the OpenAI and Claude Agents SDKs, or a custom agent defined in YAML with one-line changes, and use a /model command to swap models mid-session. Control: stateful, contextual policies that enforce guardrails at the meta-harness layer rather than via a prompt, the examples given are pausing an agent after it spends $100, or requiring human approval before a git push that follows an npm install, with YAML policies stacked across server, agent, and session scopes, strictest checked first. Collaboration: share a live agent session by URL so teammates can watch, comment, co-drive, or fork the conversation in real time. Two example agents ship in the repo: Polly, a multi-agent coding orchestrator that delegates to parallel sub-agents in git worktrees and routes each diff to a different vendor for cross-review before merging, and Debby, a brainstorming partner running both Claude and GPT that you can tell to /debate so the two critique each other before converging.
The part most worth dwelling on is where the guardrails live, because it is the answer the whole industry has been circling. Omnigent enforces policy outside the model, at the orchestration layer, not inside a system prompt the model could be argued out of, and its Omnibox OS sandbox goes further by keeping secrets out of the agent's reach entirely. A GitHub token, in their example, is never shown to the agent at all; it is injected only by the egress proxy, and only on requests the policy has approved. Set that against this week's OWASP report, which named prompt injection the top production failure mode and prescribed bounding the blast radius rather than trusting the model to police itself. Omnigent is close to a direct implementation of that prescription: assume the agent can be compromised, so limit what it can reach, gate the irreversible actions, and keep the credential somewhere neither the agent nor an injected instruction can ever see it.
For the runtime contest we keep tracking, this is the layer going meta. The fight so far has been over which harness or runtime wins, OpenClaw, Hermes, Claude Code, the hosted enterprise platforms. Omnigent's bet is that none of them wins outright, that engineers keep running several at once, and that the durable position is the control plane above all of them: vendor-neutral composition, cross-cutting policy, and shared sessions. It is also a quietly multi-vendor design by default, Polly cross-reviewing diffs across vendors, Debby pitting Claude against GPT, treating model diversity as a feature rather than a fallback. The honest caveats are stated plainly: it is alpha and early, setup wants Python 3.12, Node.js 22, and tmux, you bring your own models, infrastructure, and spending, and, notably for a tool whose entire pitch is interoperability, there is no MCP support yet, with the Omnigent Server MCP listed on the roadmap rather than shipped. But the idea is the kind that tends to stick: once you are juggling five agents, what you want next is not a sixth, it is one place to compose, govern, and watch them.