w3bsTeach an agent
by showing it.
Do a desktop task a few times, review what was captured, and train a small model that suggests the next click. Liquid neural networks (CfC) and a GRU baseline, trained on your machine by default. Off until you turn it on, and nothing is sent or posted without you.
curl -fsSL https://w3bs.org/training/install.sh | shw3bs training openOpen the Teach tab. To train, install the Infernet CLI and its runtime once: npm i -g @infernetprotocol/cli && infernet model runtime setup --yes.
Four steps.
Allow, then show it
Turn on recording in the Teach tab, pick one app and one task, and do the task. Passwords, PINs, one-time codes, tokens and card fields are masked as they are captured; the strict profile keeps no typed text at all.
Review
Label each run (goal, success, risk), redact anything left, drop duplicates. Nothing becomes training data until you mark it reviewed and turn on training data.
Train
Train on your own machine (default) with Infernet: a CfC liquid neural network and a GRU baseline on the same splits. The liquid model is chosen only if it beats the baseline. A trusted c0mpute worker is opt-in, behind a preview of exactly what leaves your machine.
Evaluate, then use
Each model is replayed on mock email, Reddit and LinkedIn sites before it can be switched on. Then pick a mode per task: Off, Suggest, Execute with approval, or Bounded. Every send, post, publish, delete or submit waits for you in every mode.
Five switches, all off.
Each permission is separate, shown with who turned it on and when, and can be revoked in one click. A kill switch stops the agent before its next action. Every proposal, approval and action is in the audit log.
- recording: capture a scoped Teach session
- training data: use reviewed episodes to train
- network upload: send a redacted dataset to a trusted worker
- remote inference: ask a remote worker for suggestions
- auto-execute: let a model act (external writes still ask you)
Open pieces, one contract.
Infernet trains and serves the models (CfC and LTC from ncps, GRU baseline). c0mpute runs them on a worker you trust when you ask it to, with end-to-end encrypted sessions. All three speak the liquid-agent/v1 schemas.