agentspan-ai/agentspan

What it solves

Agentspan provides a durable runtime for AI agents, ensuring that agent executions do not fail or disappear when a worker process crashes or a system restarts. It solves the problem of "in-memory" agent state, allowing agents to run for minutes, hours, or even days—especially when waiting for human approval—without losing progress.

How it works

Built on Conductor, Agentspan compiles agent definitions into server-side executions. This moves the orchestration logic from the client-side loop to the server, enabling automatic resume from the last completed step after a crash. It supports a "Plan-Execute" strategy where an LLM generates a JSON plan that is then executed deterministically. It also integrates with various event sources (cron, Kafka, SQS, webhooks) to trigger agents.

Who it’s for

Developers building production-grade AI agents that require high reliability, distributed tool execution across different languages, and durable human-in-the-loop workflows.

Highlights

• Durable Execution: Automatic crash recovery and the ability to resume agents from the last completed step.
• Human-in-the-Loop: Tools can be marked as requiring approval, pausing execution durably until a human intervenes from any machine.
• Distributed Workers: Tools can be executed as distributed tasks in Python, Java, Go, or C#.
• Server-Side Tools: Direct integration with OpenAPI/Swagger specs, HTTP endpoints, and MCP servers without needing separate workers.
• Framework Compatible: Works with existing agents from LangGraph, OpenAI Agents SDK, and Google ADK.
• Production Guardrails: Support for custom functions, regex, or LLM judges with configurable failure modes (retry, raise, fix, or escalate).
• Secure Credential Management: Encrypted server-side storage for API keys and secrets, injected into tools at runtime.