Own worktree
Every agent works on its own git branch.
Getting things ready…
Many agents. One project. You stay in charge.
I'm still learning. I wanted agents to build quickly without leaving me with code I can't read or habits I didn't choose.
The app enforces it through each agent's own permission flags, not by asking nicely.
Reason together. You put your approach first.


The same question gets a different answer depending on what you already know.
YouAdd a job queue to the server.
AgentThree ways to do this: in-memory array (cheapest, loses jobs on restart), a JSON file written atomically (survives restarts, no dependency), or SQLite (real queries, new dependency). I'd pick the file for this size. What's your instinct?
The rule it is given: Two or three options with complexity and trade-offs, one recommendation, and only the new concepts explained.
Illustrative wording for the rule above, not a recorded transcript.

Learning TODO mode: the agent handles error handling and plumbing, and leaves the decisions as tasks you can click straight into.
Decide when to retry. Network blips yes, bad input no. What is your rule?

Every agent works on its own git branch.
See each change, with a note, before it lands.
Ctrl+` in the repo or any worktree.
tandem/ branches only. Never forced.
Runs locally. Chats stay on this machine. Agent coordination and release polish still need work.
STACK
Measured from the project's own source: 149 files.
Libraries & tools