Introduction

Agent First Organization provides a framework for developing task-oriented dialogue agents to complete complex tasks powered by LLMs. The framework is designed to be modular and extensible, allowing developers to customize workers that can interact with each other in a variety of ways under the supervision of the orchestrator managed by Taskgraph. The framework is built on top of the LangChain platform, which provides a flexible LangGraph framework and LangChain Expression Language (LCEL) to chain different module together.

In this framework, we propose opportunities for innovation across four key areas:

• Mixed control: The workers are designed to handle different goals from user needs and builder objectives, ensuring more dynamic and goal-driven interactions. Use Config to specify different goals and constraints.
• Task Composition: The complex real-world tasks are proposed to be broken down into modular, reusable components handled by each worker, enhancing efficiency and scalability. The Taskgraph is used to handle the task composition.
• Human Intervention: Integrates human oversight and interactive refinement, ensuring critical decisions are accurate and user preferences are prioritized. Customize Worker to incorporate human intervention.
• Continual Learning: The workers evolve and improve over time by learning from interactions, maintaining their relevance and effectiveness in dynamic environments. The Evaluation will be used to analyze the performance and further improve the Taskgraph and Workers.

They focus on improving worker's controllability, reliability, and transparency, fostering a new generation of AI Agents capable of dynamic collaboration and self-improvement.