Not now Start Free Trial

How many times auto-heal will restart this agent per hour. Prevents infinite loops.

⬆️ Increase (5-10): ✅ Critical agents stay up through transient crashes. ❌ More restarts = more API cost (each restart re-loads the full system prompt).
⬇️ Decrease (1-2): ✅ Forces you to notice real problems faster — agent stays down, you get alerted. ✅ Lower resource waste on permanently broken agents. ❌ One-off transient crashes won't self-recover — you'll be woken up for noise. ❌ Dev/staging agents that crash frequently will stay down.
🟢 Rule of thumb: Production agents → 5. Dev agents → 1. Unknown → keep at 3.

System prompt growth % that triggers a warning over 7 days.

⬆️ Increase (20-30%): ✅ Fewer false positives — only flags significant bloat. ✅ Good for actively developed agents (expected growth from new skills). ❌ May miss gradual token creep that silently adds $5-10/mo over months.
⬇️ Decrease (5-10%): ✅ Catches bloat early — every small composition change gets reviewed. ✅ Keeps token costs predictable. ❌ Will flag during active development even if changes are intentional. ❌ More notifications to review.
🟢 Rule of thumb: Stable production agents → 10%. Actively developed → 25%. Don't-know → 15%.

Accumulated memory tokens before cleanup triggers.

⬆️ Increase (80-120): ✅ Richer historical awareness — agent remembers more between sessions. ✅ Less frequent garden cleanups (disruptive if mid-conversation). ❌ Higher per-turn token cost. ❌ Memory debt can compound — duplicates and contradictions accumulate before cleanup fires.
⬇️ Decrease (20-40): ✅ Keeps memory lean — lower token cost per turn. ✅ Catches contradictions and duplicates early before they compound. ❌ May trim useful context too aggressively — agent loses conversational threads. ❌ More frequent garden runs = more background processing.
🟢 Rule of thumb: Long-running agents with complex memory → 80-100. Simple stateless agents → 30. Not sure → 60.