LANGCHAIN — BCG-X Releasing AgentKit, a Full-Stack Starter Kit for Building Constrained Agents

Artificial intelligence is growing up fast, as are robots whose facial expressions can elicit empathy and make your mirror neurons quiver. — Diane Ackerman

BCG X has released AgentKit, a full-stack starter kit for building constrained agent applications. This article will provide you with an overview of AgentKit and how it works, as well as the motivation behind its development.

Introduction

AgentKit is a starter kit for building full-stack constrained agent applications. It is built on the latest versions of Next.js 14, FastAPI (Pydantic 2.x), and LangChain for optimal performance, security, and developer experience.

What Is AgentKit?

With AgentKit, developers can quickly experiment with a constrained agent architecture with a chat UI and scalable backend, and build a full-stack, chat-based agent app that can scale to MVP.

How it works

At the heart of the AgentKit framework lies the concept of “Action Plans”, a predefined tree of tasks and decisions that the agent navigates to get to the desired output. Here’s a simplified example of how AgentKit works:

# Selecting an Action Plan
selected_action_plan = router_agent.select_action_plan(user_prompt, chat_history)

# Executing Action Steps
for action_step in selected_action_plan.steps:
    # Execute tools asynchronously
    output = action_step.execute_tools(context)
    # Signal status updates and stream output
    signal_status_update(output)

The code above illustrates a simplified version of how AgentKit selects an Action Plan and executes its corresponding Action Steps.

Motivation behind AgentKit

AgentKit was developed to prioritize speed-to-value, rapid development, scalability, reliability, and user-centric design. It aims to enable developers to build an initial proof of concept fast, then iterate with users.

Build an initial proof of concept fast, then iterate with users

With the configurability and modularity of the AgentKit tech stack combined with the flexibility of LangChain, a strong first proof of concept can be built in just a few days.

Make applications that easily scale to MVP

AgentKit provides pre-built modules for authentication, user feedback monitoring, queue management, and caching, making it easy to deploy and scale applications to production.

Develop applications that are safe and reliable

LangSmith has been a useful tool for evaluating the reliability and safety of agent outputs, ensuring “cognitive control” over the agent’s actions.

Build applications for business users

AgentKit’s UI is designed to show what the agent is doing at every step, using streaming to display its preliminary outputs. It also includes support for rendering tables, dynamic visualizations, and code to support a wide range of business use cases.

This article has provided an overview of BCG X’s AgentKit, a full-stack starter kit for building constrained agent applications, delving into its functionality and the motivation behind its development. If you are interested in learning more or trying out AgentKit for yourself, you can explore the GitHub repository or AgentKit documentation.