What does Ronin actually do to my Agent?
Nothing inside. Everything at the edges.
Ronin does not alter your Agent's own capabilities. It shapes one thing: how an Agent meets other Agents and the house. Where work starts, where it goes when done, what the Agent is told along the way. That is coworking. If you read one page about what Ronin does, read this one.
The picture
Your Agent's world is its own. Ronin stands at the doors.
An Agent under Ronin is the Agent you would run alone: same LLM, account, editor and terminal. Ronin never reaches inside. It puts a few doors on the boundary, each a plain command, and says what they are for.
How Ronin shapes an Agent
Three ways, and only three.
- Describe the method, at the edges.Where work starts, where it goes when done, what to check when surprised. Never how to do the job. For a repository the whole method is three verbs: get a worktree, update it, hand it in.
- Expose edge tools.One verb, one command, one job. The plain call is the clean path. A tool never does a second job after its first; a second job worth doing is a second tool.
- Provide information.Every tool answers with an ack: what it did, the facts Ronin already holds, and, when the Agent should act, a nudge naming the tool. Nothing in an ack runs anything.
Independent by default
Start bare. Add only the coordination you want.
With nothing switched on, an Agent is a plain Agent in a plain checkout. Messaging, the team board, worktrees and a Team lead are separate routines. Each adds one door and the sentence that explains it. Turn it off and the door is gone.
A routine is an offered path, not a lock. Ronin hands the Agent a command; it does not reach past the door and make it do anything.
Coworking, defined
What coworking is, and what it is not.
| Coworking iscoordination at the edges | Coworking is notcapability inside | |
|---|---|---|
| LLM | Whatever provider and account you chose. Ronin runs the CLI you already use. | An LLM, a wrapper, a prompt library, or a "smarter" Agent. |
| Work | Where work starts and where it lands: a private worktree, a hand-in, a lead's review. | How the Agent writes code, tests, plans, or reasons. |
| Tools | A handful of edge commands: get, update, hand in, status, close. | Tools for the job itself. Your Agent brings its own. |
| Words | Acks with facts and named nudges. Warnings where something could be lost. | Gates, refusals, or actions taken for the Agent because it "should" have. |
| Memory | Receipts: what was handed in, promoted, closed, by whom, when. | Persistent memory for the Agent, or a record of what it thought. |
| Publishing | A pull request you review and merge. | Anything that pushes to your remote by itself. |
If anything in the right-hand column became true, Ronin would be doing the work instead of coordinating it.
Information, not automation
What an ack looks like.
The reply after a hand-in. Each line is what happened, a fact Ronin already held, or a nudge naming a tool.
ACCEPTED ronin_cowork:team/alpha/maya → team/alpha/dev now 4f2a9c1e07 [hi_20260904104839_7f9d24] desk is level with the line base 51ce9424 · 0 behind dev · line is 3 ahead of dev; promotion due unsaved in this desk: none your lead is kenji NEXT: line moved; run tejun-desk status; if it reports a dev update, run tejun-desk sync NEXT: tell your lead with tejun-send kenji "..."
Nothing ran on the Agent's behalf. Each follow-on is one command away, and the ack said which.