Agentspan

Agentspan is what happens when developers finally admit that most AI agent frameworks are glorified wrappers with zero fault tolerance. Your agent starts running, your process OOMs at step 3, and poof—there goes four minutes of LLM calls and expensive API credits. Agentspan fixes this. It's an MIT-licensed, self-hostable runtime that keeps your agent alive even when your worker dies. Execution state lives on the server. You reconnect from any machine, any time, and it resumes from exactly where it stopped. No lost work. No retries from scratch.

But the real magic is human-in-the-loop. One decorator marks any tool as requiring approval—the agent pauses, holds state with no timeout, and waits. Approve from Slack, a webhook, or code. Process refunds, flag suspicious transactions, gate sensitive actions—your AI can now wait for humans precisely when it matters, not before or after. And pipeline mode lets you wire multiple agents together in one expression, where each output feeds the next, every step logged and crash-safe across the full chain.

This is the infrastructure layer nobody's building but everyone's needing. As agents get more ambitious and run longer, process death isn't a "if"—it's a "when." Agentspan just made that someone's else's problem.