Orkes Unveils Agentspan: A New Standard for Durable AI Agents in Production

Orkes has introduced Agentspan, an open-source runtime environment engineered to tackle the inherent fragility of AI agents in production. While AI agents move beyond traditional chatbots to autonomously achieve complex goals—such as searching, comparing, and booking—their deployment often faces significant hurdles. Common frameworks frequently result in process crashes, loss of execution state, and a critical lack of visibility, rendering agents unreliable beyond local testing. Agentspan addresses these issues by compiling agent code into a durable workflow, detaching the execution state from the local process and hosting it on a dedicated server. This architectural shift guarantees agent survivability through crashes, preserves progress, and provides comprehensive insights into every step of the execution lifecycle, including the ability to pause and resume operations mid-run.

The platform’s efficacy was demonstrated through a Flask-based AI concept explainer web app, leveraging a local Llama 3.1 model via Olama. This setup illustrated Agentspan’s capability to track an agent’s full execution history—including LLM calls, tool executions, and timing—through its intuitive UI at localhost:6767. A critical test of durability involved simulating an agent crash; Agentspan successfully preserved the agent’s state on the server, allowing the workflow to be seamlessly resumed and completed without any loss of progress or manual state restoration. Furthermore, Agentspan supports integration with agents built using other prominent frameworks like LangChain, Google ADK, and OpenAI Agents SDK, positioning it as a versatile solution for developers building robust and resilient AI agent applications for production.