Set three nested timeouts on autonomous agents: per-tool-call, per-task-loop, and per-sandbox-lifetime

Firecrawl·medium signal

Bound runaway agents with timeouts at three distinct levels — per tool call (~30s), per task loop (~20min), and per sandbox lifetime — so a single hung tool, a looping reasoning chain, and a zombie sandbox each have their own independent kill point. Most teams set only a single overall timeout and discover too late that an agent burned tokens looping inside a still-'alive' session. Layering the limits gives containment at the exact granularity where each failure mode actually occurs.