Two public cases got me building this: $1,800 in two nights of overnight Claude Code runs (issue #37686) and a $437 loop on one missing file. Both looked green in the logs.

Hi PH — I ran autonomous agents and a crypto trading bot unattended for two years. Every expensive failure was quiet: a routine that "succeeded" with an empty report, a loop that called the same tool 41 times, a bill that doubled overnight while the dashboard lagged a day.

Ping monitors (Healthchecks, Cronitor) know a job ran. Tracing platforms (Langfuse, LangSmith) need an SDK and a team. RunVouch is the missing middle for people who run agents unattended:

• MISSED / STALLED — the clock lives outside the agent, so silence is an alert
• NO_EVIDENCE — a run only counts as done when a file changed, a URL returns 200, or your assertion passes. Anthropic's own docs say green ≠ succeeded.
• RETRY_STORM — same tool + identical input 8× in one run
• BUDGET — per-run and per-day cost caps, read from Claude Code transcripts
• DRIFT — output size / duration off its 7-run baseline

Two lines around any job (rv run NAME -- your-command), a Claude Code plugin, Python/Node clients, and an MCP server in the official registry so your agents can check on each other. Fails open: if RunVouch is down, your job still runs. Free for 3 agents, no card; MIT self-host.

Which failure mode am I missing? Genuinely want to know.

Two public cases got me building this: $1,800 in two nights of overnight Claude Code runs (issue #37686) and a $437 loop on one missing file. Both looked green in the logs.

RunVouch

The watchdog for unattended AI agents

The watchdog for unattended AI agents