keepalive

keepalive watches your training runs and fixes them while you sleep. You write a plain-English sentence describing what healthy training looks like; we run a monitoring agent against your live metrics, and when something drifts, a fleet of coding agents open pull requests with candidate fixes — each with a full report.

One-liner: agents that hold you accountable — and stop waiting when you don't show up.

What you get

• A monitoring agent that reads your metrics every few seconds and scores them against your prompt: "Flag if val/loss diverges from train/loss or grad norms spike."
• A fixing pipeline that, on an incident, spawns up to max_agents coding agents. Each works in its own sandbox on its own branch and opens a PR with a written report.
• Training that launches itself. Merge to your default branch and the run starts in a W&B Sandbox, reporting metrics straight back to the dashboard.
• Full traceability. Every agent decision is one Weave trace tree per incident.

The flow

  merge to main
        │
        ▼
  ┌──────────────────────┐
  │  W&B Sandbox training │  ── metrics ──▶  POST /api/v1/events
  └──────────────────────┘
        │
        ▼
  ┌──────────────────────┐
  │   monitoring agent    │  scores each window  →  { confidence, reasoning }
  └──────────────────────┘
        │  confidence < threshold  (or a hard failure: NaN / OOM / crash)
        ▼
  ┌──────────────────────┐
  │    fixing pipeline    │  hypothesis agent → N coding agents (one per sandbox)
  └──────────────────────┘
        │
        ▼
   N pull requests, each with a hypothesis + a full report

Get going

• Quickstart — sign in, create a key, wire up your loop, merge.
• Agent blurb — the prompt you hand your coding agent to wire keepalive in.
• Concepts — how the monitoring and fixing agents actually work.
• Reference — keepalive.watchdog params, env vars, CLI, events.