Dev instincts.Designer eyes.No sacred rules.

My unattended agent runs kept dying silently, and it took me too long to realize a permission prompt was the culprit: the process waits forever instead of failing loudly. Now I treat every headless run as needing an explicit permission mode, set up front, before I trust it in CI. If a tool can block on a human, that's a hang in any script, so I design for that assumption from the start.

Retry queues sound safe until surge traffic turns them into a deadline killer. When AI calls hit provider rate limits, the queue silently swallows hours you budgeted for debugging. I now plan for burst capacity up front, or shed load deliberately, instead of trusting retries to absorb the spike.

I lost an hour of agent reasoning the moment OpenCode restarted, and realized I'd been treating the terminal as durable storage when it never was. Scrollback is context too: without it, I can't audit what the agent actually did versus what it claims. Now I check whether a tool persists history before I trust it with long tasks.