LangGraph and the OpenAI Agents SDK are typical choices for this part of a deployed agent, the code that executes its loop. Around it sit a wrapping server and a hosting runtime.
Also called agent loop.
Read more: Databricks docs
In the Ultra Transcenders books
Each book explains Framework or harness in context, with comparison tables and the common traps.
Terms in this definition
- LangGraph
Build tool-calling loops, or apps where an LLM works with state and multiple actors, using this open-source library. A ready-made template comes with the Agent Bricks CLI.
- OpenAI Agents SDK
OpenAI's framework for multi-agent work, with an agent runner, sessions and tools, and no reliance on LangChain. Projects for it can be scaffolded with the Agent Bricks CLI.
- Agent
A specialised form of Microsoft Copilot set up for one particular job, pairing instructions with knowledge and skills. You can create one in Copilot Studio, SharePoint or Agent Builder, and administrators control them from the Microsoft 365 admin center.
Related terms
- Agent server
A library that puts an HTTP server around an agent loop and sets the behaviour for runs that get interrupted. For new agents, the recommended choice is
DurableAgentServerfrom AgentKit. - DurableAgentServer
The recommended server for new code agents: a class in AgentKit's
databricks_agentkitmodule that hosts an agent loop over HTTP with the invocation API, follows every run and recovers those that a crash cut short.