I’m a little tired of elaborate workflows to avoid contaminating coding agents’ context with ambiguous discussions that end up dissolving the specs they produce.
So I did what anyone would do: I built “Corporal Code” ( GitHub - JanBanholzer/corporal-code · GitHub ) as an extension for Cursor that keeps a context and UI-aware, as well as tool-capable Gemini 3.8 Live conversation always-on while you work. It hears what you say and, at two screenshots a minute by default, sees what you do, enough to notice when you’ve started editing something beyond tab-complete and suggest delegating it. (only on Mac right now)
The goal is to separate conversation from implementation while maintaining supervision. You’re the Sergeant, “Code” is your Corporal and the Cursor agents are the privates you send running around. The analogy works because it will still be chaos.
You discuss ideas, plans, tradeoffs, and implementation details in an uninterrupted conversation, and Gemini delegates implementation tasks to Cursor agents on instruction.
If you’re scaffolding UI with a device simulator or on localhost, it’s already great: set the config to delegate background agents immediately per default, use a low latency model and increase the screenshot rate - then you can go hands free and just chat & experiment.
The most interesting part from my perspective isn’t hands-free, voice input, screen capture (like that’s new) or separating discussion and delegation from implementation while keeping a conversation alive. It’s that now you can do all of those at once while ALSO watching the agents’ diffs in real time the Cursor way, not just when they land.
In the default config, it uses extended thinking for Gemini. It drafts a prompt and opens a new Cursor chat for you to review before anything is executed. You can edit the prompt, send it, watch the agent work, and review the diffs as usual. I’ve uploaded a short demo of that workflow to the readme on gh.
I’d love to hear whether this workflow feels like an upgrade to anyone else!
There are still caveats as the Cursor SDK doesn’t interact very well with Gemini’s limitations, especially on context handling, but the basic approach might be worth building on.
Unfortunately, the SDK doesn’t expose a path for native agents themselves to be visible to Gemini (background & cloud are - but not in the foreground). You can however supervise the implementation directly instead of waiting for an entire workflow to land just to be reverted, when you could have stopped the agent from stepping on “that landmine again” during the first few edits.
To mitigate that agent-exposure limitation, you can change the code to send a few screenshots per second to Gemini, that way it’ll see what you’re seeing. And trigger reasoning cascades that hurt your wallet if you’re my size.
You shouldn’t mind the Google API bill going from the $3/h I’ve measured in testing up to anywhere around the fifties doing that - so get rich, use a company card or turn it off during pauses, keeping your sessions short.