One agent, many providers
Claude Code can plan a change and hand the test to Codex, the review to Cursor and the docs to OpenCode. Each helper runs in its own thread with its provider's own session, history and tools, signed in through that CLI, so you can open any of them and keep talking to it.
Agents do it with ace's MCP tools, the same in every CLI: delegate starts helpers with a task, wait collects their results as they finish, agents_status shows where each one stands, and agent_message and agent_cancel steer or stop one.
Send work to the right machine
A helper can run on another of your linked machines: the long soak test on the Linux build server, the Windows build on the workstation, the iOS check on the Mac with the Simulator. When that machine is busy, the work waits in its queue instead of failing.
Status you can trust
A thread's status covers its whole tree: every helper, the helpers they started in turn, and their background commands. ace never shows a thread as done while any agent below it is still working or waiting on you, and a question from a helper several levels down shows on the thread you are looking at.
Stop stops the whole tree. A restart doesn't lose work: results that arrive while ace is down are delivered when it comes back.
Limits you set
A thread can run up to 64 helpers, up to 8 levels deep. In Settings you choose how many agents a thread runs at once, and the rest wait their turn. A helper never gets more access than the agent that started it: it inherits the parent's permission mode and can only lower it.