Agentspan

Agentspan is an open-source server and SDK designed to provide durable execution for AI agents. It compiles agent definitions into robust workflows, ensuring that agent execution state persists outside of the immediate process. This means agents can automatically retry tool calls, gracefully handle process crashes (like OOM errors or deployments), and resume from the exact point of failure on any machine, preventing lost work. The platform is built for production environments, addressing critical needs like crash recovery, human-in-the-loop interactions, and complex agent pipelines. It leverages Conductor, an open-source orchestration engine used by major tech companies, to provide primitives like durable state, per-step retries, full execution history, and replayability. Agentspan offers a clean Python API on top of this foundation, making it accessible for developers building and deploying AI agents. It also includes comprehensive observability tools to monitor every step, tool call, and LLM request, along with robust testing capabilities that allow for deterministic testing of agent logic without needing an LLM or server. Agentspan is ideal for developers and teams looking to deploy reliable, production-grade AI agents that can withstand failures, integrate human oversight, and orchestrate complex multi-agent workflows. Its self-hostable nature and open-source foundation provide flexibility and control over the agent infrastructure.