{"slug": "databricks-omnigent-one-control-plane-for-every-ai-agent-you-run", "title": "Databricks Omnigent: One Control Plane for Every AI Agent You Run", "summary": "Databricks open-sourced Omnigent, a meta-harness that provides a single control plane for governance, cost, and cross-device collaboration across existing AI agent runtimes, on June 16 under Apache 2.0, already garnering ~7,500 GitHub stars. Omnigent enforces policy at the infrastructure layer—such as spend caps with escalation and human-in-the-loop for risky actions—rather than via system prompts, and allows swapping harnesses like Claude SDK or Codex with a single config line. Sessions sync across devices and can be shared with teammates for live collaboration.", "body_md": "Most developers running multiple AI agents in 2026 are quietly managing three separate billing streams, writing governance rules into system prompts the model can silently ignore, and losing all session context every time they switch between Claude Code and Codex. Databricks just open-sourced a direct fix: **Omnigent**, a meta-harness that sits above all your existing agent runtimes and gives you one control plane for governance, cost, and cross-device collaboration. It landed on June 16 under Apache 2.0, and it already has ~7,500 [GitHub stars](https://github.com/omnigent-ai/omnigent).\n\n## A Meta-Harness Is Not Another Framework\n\nBefore anything else: Omnigent is not a replacement for Claude Code, Codex, Cursor, LangGraph, or CrewAI. It is a layer *above* them. You define your agents in a short YAML file and point Omnigent at the harness you want to run. Your prompts, tools, and sub-agents stay the same. The harness is just a config line.\n\nThis distinction matters. The existing frameworks — LangGraph, CrewAI, AutoGen — ask you to write agents *inside* their abstractions. Omnigent asks you to write agents *for* whatever runtime you already use, then adds orchestration and governance on top. Your LangGraph agent can live inside the `tools:`\n\nblock of an Omnigent YAML. They are not competing.\n\n## Policy Enforcement That the Model Cannot Override\n\nThis is the design decision that makes Omnigent worth paying attention to. Governance via system prompt is not governance — it is a suggestion. Any sufficiently long session can drift past it. Omnigent moves policy enforcement to the infrastructure layer, where the model’s cooperation is not a prerequisite.\n\nSome examples directly from the [Databricks announcement](https://www.databricks.com/blog/introducing-omnigent-meta-harness-combine-control-and-share-your-agents):\n\n**Spend caps with escalation:** Soft warning at $3.00; hard block at $5.00 per session. The agent does not decide when it has spent enough.**Human-in-the-loop for risky actions:** Require approval before`git push`\n\nif the agent installed a new npm package during the session.**Contextual policies**(added in July 2026): Policies can inspect session history. “If the agent has made more than three file writes in the last ten minutes, pause and ask.” Stateful enforcement, not just a blanket rule.**Filesystem sandboxing:** Read-only access outside the project directory and`/tmp`\n\n.\n\nNone of these require model cooperation. They run at the harness layer and either allow, block, or pause the next action. For teams that have watched an agentic session blow past its intended scope, this matters.\n\n## Swapping Harnesses Is One Line\n\nThe practical benefit developers are posting about most is the harness swap. An Omnigent agent definition looks like this:\n\n```\nname: my_dev_agent\nprompt: You are a senior backend engineer. Focus on correctness.\nexecutor:\n  type: omnigent\n  config:\n    harness: claude-sdk\n    model: claude-sonnet-4-6\ntools:\n  - type: bash\n  - type: mcp_server\n    config:\n      command: npx -y @modelcontextprotocol/server-filesystem\n```\n\nChange `harness: claude-sdk`\n\nto `harness: codex`\n\nand the same agent runs on Codex. Change it to `cursor-native`\n\nand it runs inside Cursor. Your prompt, tools, and policies carry over unchanged. Setup auto-detects credentials already on your machine — if the Claude CLI and Codex CLI are installed and authenticated, Omnigent picks them up as subscription-backed harnesses. No new API keys, no additional billing.\n\n## Sessions That Follow You Across Devices\n\nOmnigent exposes every session over a terminal, a local web UI at `localhost:6767`\n\n, a macOS native app, and mobile browsers. The tabs — Files, Agents, Shells, Todos — sync in real time. Click Share on any session and send the URL; a teammate opens it in their browser and joins the live session, with the full agent context intact. They can watch output stream, send messages, comment on specific outputs, or fork the session to try a different direction.\n\n## What to Know Before You Install It\n\nOmnigent is in alpha. The honest picture:\n\n**Windows support is degraded.** The server and web UI work, but native PTY/tmux wrappers and stronger sandboxing are not yet available. Targeted for Q3 2026.**Not production-ready for critical workloads.** Some users report dependency errors on install. Evaluate this in a test environment, not against live systems.**Watch the managed version.** Omnigent on Databricks integrates with Unity AI Gateway and Unity Catalog for centralized cost tracking, audit logs, and smart routing. The enterprise pitch is real — and so is the pull toward the Databricks ecosystem. The Apache 2.0 license means you can fork and self-host, but know which version you are evaluating.\n\nInstall is one command:\n\n```\ncurl -fsSL https://omnigent.ai/install.sh | sh\nomnigent setup\nomnigent\n```\n\nThe setup command scans for existing credentials and offers detected harnesses as defaults. The web UI spins up at `localhost:6767`\n\nautomatically. Full [harness documentation](https://omnigent.ai/docs/build/harnesses) is available on the Omnigent site, and [Help Net Security has a solid third-party breakdown](https://www.helpnetsecurity.com/2026/07/06/omnigent-open-source-ai-agent-framework/) if you want an independent read before committing.\n\n## The Bottom Line\n\nIf you are running more than one AI agent harness, you need a governance layer above them. Prompts are not that layer. Omnigent is the first credible open-source attempt to build it correctly — policies at the infrastructure level, harness portability that actually works, and cross-device sessions that preserve context. It is early, and the rough edges are real. But the design is sound, the license is right, and the alternative — multiple agent setups with separate policies that each model can quietly ignore — is worse. Install it in a test environment today.", "url": "https://wpnews.pro/news/databricks-omnigent-one-control-plane-for-every-ai-agent-you-run", "canonical_source": "https://byteiota.com/databricks-omnigent-meta-harness/", "published_at": "2026-07-23 03:10:32+00:00", "updated_at": "2026-07-23 03:31:47.114644+00:00", "lang": "en", "topics": ["ai-agents", "ai-infrastructure", "ai-tools", "developer-tools", "ai-policy"], "entities": ["Databricks", "Omnigent", "Claude Code", "Codex", "Cursor", "LangGraph", "CrewAI", "AutoGen"], "alternates": {"html": "https://wpnews.pro/news/databricks-omnigent-one-control-plane-for-every-ai-agent-you-run", "markdown": "https://wpnews.pro/news/databricks-omnigent-one-control-plane-for-every-ai-agent-you-run.md", "text": "https://wpnews.pro/news/databricks-omnigent-one-control-plane-for-every-ai-agent-you-run.txt", "jsonld": "https://wpnews.pro/news/databricks-omnigent-one-control-plane-for-every-ai-agent-you-run.jsonld"}}