Skip to main content
The team watchdog is a safety feature that detects stalled teammate agents and restarts them. It is designed to be conservative, configurable, and to prevent runaway loops that could balloon LLM costs.

​
What it monitors

The watchdog monitors teammate activity (for example: work starting, tool calls, model calls, and errors) and looks for signs that a teammate has stopped making progress.

​
When it runs

The watchdog runs periodically in the background.

​
Staleness detection

Staleness detection uses conservative time thresholds and includes guardrails to avoid restarting brand-new teammates or thrashing during transient failures.

​
What respawn does

When a teammate is restarted, the gateway attempts to preserve continuity so another teammate can safely continue based on local state and outputs.

​
Rate limiting (prevents runaway respawn loops)

In addition to a cooldown, the watchdog enforces a cap on restarts within a time window. If a teammate hits the cap, the watchdog pauses automatic restarts to avoid runaway loops and force human intervention.

​
Configuration

These settings live under gateway.team_watchdog:
  • enabled: enable/disable the watchdog.
  • staleness thresholds: how long a teammate can be inactive before being considered stalled.
  • respawn limits: cooldowns and caps to prevent thrash.
For most setups:
  • Keep tool_stale_after_seconds higher than stale_after_seconds (tools can legitimately take longer).
  • Keep a non-zero startup_grace_seconds (spawns can be slow).
  • Keep max_respawns_per_window low enough to prevent cost runaway, but high enough to recover from transient failures.