Sign in to pick up where your crew left off.
LearnLoop
LearnLoop runs your code before it says anything about it, and the crew stays quiet unless one of them can actually help. Pick a room to start.
Every message runs the same five steps, and each one is written to a log you can read back.
Shown in every group you join. Without a photo you get a creature, picked from your name so it is always the same one.
In plain English, built from every session you have had. This is a belief, not a fact: rewrite it and the agents treat your version as more authoritative than their own estimates, in every chat and group from then on.
One line per past session, distilled from the trace with no model call. This is what the agents are shown about your history.
Loading…
Which ones you have written something in. The studio opens new editors in the first one you pick, and the agents know not to assume the others.
This is the open learner model, the AI's live beliefs about your skills. It learns from your runs, hints, and corrections. Disagree? Drag a slider: your edit is a first-class correction and the studio adapts.
One number per agent, from 0 to 1: how much that agent currently trusts its own judgement when it talks to you. Every agent starts at 0.70. Each time you correct one, its number drops by 0.10 and the correction is kept below as a blind spot in your own words.
It is here because it is the other half of the open learner model. That page shows what the studio thinks of you; this shows what it admits it gets wrong about you, which is the part that is usually hidden.
Two numbers, 0 to 1, measuring different things.
Perceived reliability: how much you trust its answers. Starts at 0.70, falls 0.05 each time you correct it.
Engagement signal: how readily you push back. Starts at 0.60, rises 0.02 when you correct it or edit your own model.
Correcting the studio moves them in opposite directions on purpose. You trusted that answer less, and you checked its work instead of taking it on faith. Both are shown so you can see what the studio thinks of you. Nothing here adapts silently.
Tell the studio what your project or assignment is, every agent will know the brief when called on.
Optional, but the Planner works backwards from this date and will tell you if the scope does not fit.
Including you. Add what each person plans to work on if you know it, or leave it blank and the Planner will ask them.
Everyone with an email address here is invited automatically when you create the project. They get a link that joins them with no code to type, and a notification if they already have an account.
Got a code from a teammate? Enter it here.
No groups yet.
Agents never interrupt. /learnloop <question> invites them,
/hint asks for a hint, or click the bubble when one raises its hand.
The mentor reads your group chat history (or your personal chats if you haven't joined a group), your profile, and the studio's open learner model - then recommends concrete directions with next steps. Nothing is invented: thin evidence is said out loud.
Each slice is that direction's share of the skills the studio has actually observed, weighted by how much each direction leans on them. It is a reading of your evidence so far, not a prediction, and it moves as the studio watches you work.
Each time you ask, the answer is kept here with the evidence it was based on.
Loading…
Rooms you created or joined. Opening one rejoins it.
Loading…
Your own chats with the studio. Opening one resumes it.
Loading…
The live personal-chat session and what it has cost so far.
The append-only trace opens from the header of the chat it belongs to: ⛓ Trace in a group session.
Toggles apply to your next session and are logged as standing dismissals.
Everything the studio remembers about you is inspectable: the open learner model lives on your profile; distilled session memory below.
LearnLoop
refs the old one. This log is what
replay.py rebuilds the whole session from with zero model calls, what
verify.check_one_speaker tests the one-speaker rule against, and what every
number in metrics.md is computed from (proposal §4.10).