The tutorials show you how to build an AI agent. Nobody shows you what happens when you leave it running.
We’ve been running TinkerClaw agents around the clock for over six weeks. Not as a demo. Not for a benchmark. In actual production, managing real messages, real files, real cron jobs, and real money. Here’s what the getting-started guides don’t mention.
It Will Crash. A Lot.
75 config crashes in the first two weeks. Not bugs — edge cases that only appear when an agent runs long enough to encounter them. Bad cron timing. Context windows that grow silently until they hit limits. Memory files that accumulate faster than you expected.
The fix isn’t preventing crashes — it’s building resilience. Auto-restart, state recovery, graceful degradation. Your agent needs to survive its own mistakes.
It Will Do Something Embarrassing
45 messages to the wrong chat over 10 days. That’s our record. The agent had access to multiple channels and occasionally confused which conversation it was in. Nobody got hurt, but imagine if those messages contained sensitive data.
Lesson: safety boundaries aren’t optional for production agents. Pre-send validation, channel confirmation, and trust tiers — all non-negotiable.
It Will Cost More Than You Think
An agent that runs 24/7 makes hundreds of API calls per day. If you’re using Opus for everything, you’re burning money on routine tasks. The solution: budget-aware model routing. Match the model to the task.
The Production Checklist
After six weeks, this is what we consider non-negotiable for a 24/7 agent:
- Auto-restart with state recovery — crashes shouldn’t lose context
- Channel safety boundaries — trust tiers, pre-send validation, confirmation for external sends
- Budget-aware model routing — flat-rate for automated work, metered only when justified
- Persistent memory — the agent must remember across restarts
- Self-improving cron jobs — the META pattern
- Real-time monitoring — token dashboards, not monthly invoices
- Wind-down routine — daily self-improvement cycle, not just a diary
- A field guide — 32 lessons your agent should know before it starts
Is It Worth It?
Absolutely. An agent that runs 24/7 does things you never would manually: overnight code reviews, continuous monitoring, self-improving reports, proactive alerts. It’s not replacing you — it’s doing the work you’d never get to.
But go in with your eyes open. Production is a different game than demos.
The tutorials show you how to start. We show you what happens next.