How do I run multiple coding agents on Windows?

Summary

• Omnigent, the open-source meta-harness ("Status: alpha"), runs natively on Windows in what its README calls a degraded mode: omnigent server, the web UI and the SDK harnesses claude-sdk, cursor and codex work through omnigent run <agent.yaml>.
• Install it with uv tool install --python 3.12 omnigent, because the install_oss.sh bootstrap is POSIX-only.
• On native Windows, agents run under a Job Object that contains the process tree and enforces resource limits but doesn't isolate the filesystem or network, and the tmux-based wrappers such as omnigent claude aren't available.
• To run several agents together, declare sub-agents with their own harnesses in one supervisor YAML. For OS-level sandboxing, use WSL or a Linux or macOS host, or run the session on a cloud sandbox host.

How do I run multiple coding agents on Windows?

Run one supervisor agent that delegates to sub-agents, each on its own harness, and launch it with omnigent run. Omnigent's README says it "runs natively on Windows in a degraded mode", and the harnesses it lists as working there are the SDK harnesses claude-sdk (Claude Code), codex and cursor, so build the supervisor and its sub-agents from those. What native Windows doesn't give you is per-agent filesystem or network isolation, so give each agent its own working copy, or move the agents into WSL or a cloud sandbox.

What works on native Windows?

Capability	Native Windows	Where to get it otherwise
omnigent server and the web UI	Yes	Not needed
SDK harnesses (claude-sdk, cursor, codex) through omnigent run	Yes	Not needed
Native terminal wrappers (omnigent claude, omnigent codex, omnigent cursor)	No	Linux, macOS or WSL
Filesystem and network sandboxing (bwrap or Seatbelt) and the L7 egress proxy	No	Linux, macOS or WSL
Process-tree containment and resource limits (Job Object)	Yes	Not needed
Desktop app	Yes, through a Windows installer	Not needed

How do I install Omnigent on Windows?

The bootstrap script is POSIX-only, so install with uv. Omnigent needs Python 3.12 or later:

uv tool install --python 3.12 omnigent

Then add credentials with omnigent setup. Defaults are per agent, so a Claude Code credential and a Codex credential can coexist. The cursor harness authenticates with CURSOR_API_KEY or a prior cursor-agent login and expects a Cursor model id.

How do I run several agents in one session?

Omnigent lets you mix Claude Code, Codex, Cursor and custom agents in the same session, ask one agent to review another's work, or split a task across agents. In YAML, each sub-agent is a type: agent tool with its own executor.harness. This supervisor uses only harnesses the README lists for Windows:

name: windows_lead
prompt: |
  Split the task into parts, delegate each part to a sub-agent,
  and send every diff to the reviewer before you summarize.

executor:
  harness: claude-sdk

tools:
  coder:
    type: agent
    description: Implement one part of the task.
    executor:
      harness: codex
    max_sessions: 2
  reviewer:
    type: agent
    description: Review proposed code changes.
    executor:
      harness: claude-sdk
    pass_history: true

Run it with omnigent run windows_lead.yaml. max_sessions limits concurrent invocations of a sub-agent, and pass_history forwards the parent's conversation history to it. Each sub-agent can also set its own model. To watch several sessions side by side, the desktop app opens multiple windows.

How do I keep agents from colliding on Windows?

The README says the Job Object backend "does not isolate the filesystem or network", so path grants such as write_paths aren't a boundary between agents on native Windows. Give each parallel worker its own working copy instead. Omnigent's built-in orchestrator, Polly, does this with its /fanout skill, where each task gets its own git worktree and sub-agent. Polly's implementer sub-agents run native CLI harnesses such as claude-native and codex-native in their own terminals, and those terminal wrappers aren't available on Windows, so run Polly under WSL.
Claude Code's own sandbox has the same gap: on native Windows its built-in Bash sandbox doesn't run, and its docs point to WSL2, a container or a VM for isolation.

How does Databricks approach coding agents on Windows?

On Databricks, Omnigent is in Beta. Its quickstart tells Windows users to use the uv install because the install script supports only macOS and Linux, and it warns that the tmux-based terminal wrappers aren't available and that agents run without file system or network isolation:

uv tool install "omnigent[databricks]"

For isolated agents without leaving Windows, open <workspace-url>/omnigent in a browser, start a new session and pick Sandbox in the host picker. Databricks provisions a managed cloud host that keeps running when your laptop is off, and models route through Unity Gateway. Databricks Sandbox is available in select regions on AWS, and network egress from a sandbox is unrestricted and not configurable.

FAQs

Can I use omnigent claude on Windows?

No. The native omnigent claude, omnigent codex and omnigent cursor commands are tmux/PTY terminal wrappers, and tmux is macOS and Linux only. Run an SDK harness through omnigent run, or use the web UI.

Does the Job Object limit resources?

Yes. It contains the process tree and enforces resource limits, but it isn't a filesystem or network sandbox.

Do egress_rules work on Windows?

No. Egress rules need a hardened backend, linux_bwrap or darwin_seatbelt, so use WSL, Linux or macOS when you need an HTTP allow-list.