agentspan-ai/agentspan

What it solves

Agentspan provides a durable runtime for AI agents, ensuring they don't fail when a process crashes or times out. It solves the problem of volatile in-memory agent execution by moving the orchestration to a server-side runtime, allowing agents to run for minutes, hours, or even days without losing progress.

How it works

Built on Conductor, Agentspan compiles agent definitions into server-side executions. It uses a "Plan-Execute" strategy where an LLM creates a JSON plan that is then executed deterministically as an immutable workflow. This removes randomness from orchestration and allows for automatic resume from the last completed step if a worker process crashes. It supports distributed workers in multiple languages (Python, TypeScript, Java, C#) and can trigger agents via cron schedules, webhooks, or message queues.

Who it’s for

Developers building production-grade AI agents who need crash recovery, long-running processes, and reliable human-in-the-loop approvals. It is also suitable for teams using other frameworks like LangGraph or OpenAI Agents SDK who want to add a durability layer to their existing agents.

Highlights

• Durable Execution: Automatic resume from the last completed step after process crashes.
• Human-in-the-Loop: Durable pauses for tool calls that require human approval, which can be granted days later from any machine.
• Plan-Execute Strategy: LLM-generated plans are compiled into immutable workflows to eliminate orchestration randomness.
• Server-Side Tools: Native support for OpenAPI/Swagger specs, HTTP endpoints, and MCP servers without needing separate workers.
• Distributed Scaling: Tools can be executed as distributed tasks across workers written in different languages.
• Production Guardrails: Built-in support for custom functions, regex, or LLM judges with four failure modes (retry, raise, fix, or human escalation).