Omnigent opens live Claude Code and Codex sessions with phone control
Databricks open-sourced Omnigent, a meta-harness that runs Claude Code, Codex, Cursor, Pi, and custom agents in one live session with a collaborative web UI. The release centralizes supervision, cost control, and cross-agent review instead of splitting work across separate tools.
TL;DR
- Yuchenj_UW's launch post says Databricks open-sourced Omnigent as a meta-harness that runs Claude Code, Codex, Cursor, Pi, and custom agents in one live session.
- Yuchenj_UW's UI post and Yuchenj_UW's reply about supervisor flows frame the core interaction as a supervisor agent coordinating multiple coding agents, including cross-review between Claude Code and Codex.
- According to Yuchenj_UW's UI post, the web app supports shared live collaboration and phone control, while Yuchenj_UW's summit post links to both the Omnigent site and the GitHub repo.
- The timing lands in the middle of a broader mobile-control push: reach_vb's Codex Mobile update list shipped deeper remote workflows, and NickADobos on Cursor mobile beta said Cursor mobile now includes remote control to a Mac host.
You can check the Omnigent site, browse the GitHub repo, and Yuchenj_UW's UI thread shows the part most people will care about first: Claude Code and Codex arguing inside one shared session. Yuchenj_UW's Antigravity reply also suggests the tool list was still changing days before the Databricks Summit talk. Meanwhile mattshumer_'s phone-control question and NickADobos's Cursor mobile post show how quickly remote agent control has become a product category instead of a one-off hack.
One live session
Omnigent's pitch is simple: stop juggling separate agent windows and put the whole team in one harness. Yuchenj_UW's launch post names Claude Code, Codex, Cursor, Pi, and user-defined agents as day-one participants.
The release was announced at Databricks Data + AI Summit, and Yuchenj_UW's summit post points readers to both the product site and the open-source repo. That makes this less of a demo thread and more of a real code drop.
Supervisor agent
Yuchenj_UW on supervisor orchestration describes the interesting part more clearly than the launch blurb does: a supervisor agent can talk to Claude Code and Codex, then ask them to review each other's code.
That gives Omnigent a specific workflow shape:
- one supervisor delegates work across agents
- subagents can produce competing answers
- agents can review each other's output
- a human can step in at the supervisor or subagent level
Yuchenj_UW's launch post calls the system a meta-harness built from internal Databricks dev tools, which is the right mental model here. The novelty is not another standalone coding agent. It is the coordination layer on top.
Web UI and phone control
According to Yuchenj_UW's UI post, the web UI supports three things in the same session: multi-agent debate, Google-Docs-style human collaboration, and phone control.
The phone piece matters because it changes what "remote coding agent" means. Instead of only watching logs or approving a diff, Yuchenj_UW on supervisor orchestration says you can talk directly to the supervisor or to subagents from your phone.
Mobile control is getting crowded
The same week, reach_vb's Codex Mobile update list rolled out a long list of Codex Mobile upgrades, including browsing workspace files, linking paths into prompts, picking workspace folders, reviewing diffs, approving MCP actions, and smoother /goal workflows.
Cursor is moving in the same direction. NickADobos on Cursor mobile beta said Cursor mobile entered beta with remote control to a Mac host, and NickADobos's availability reply added that it is beta-only for now.
The demand signal was already obvious before Omnigent shipped. In mattshumer_'s phone-control question, Matt Shumer asked for a way to control Claude Code on a computer from a phone with dangerous permissions, explicitly comparing the idea to Codex Mobile.