AutoForge

Inspiration

It is not hard for a software developer to build an agent using an open-source framework; however, it could be hard for a marketer.

We wanted to build a personal assistant beyond static, single-agent responses. Inspired by the potential of multi-agent systems, we aimed to create a tool that could dynamically generate and scale specialized agents for users who can't code, making automation as versatile and adaptive as a human team.

What it does

AutoForge is an intelligent assistant with multi-agent capabilities. It dynamically generates and orchestrates new agents to handle complex, multi-step tasks. It’s a flexible system that can create specialized agents for scheduling, data analysis, communication management, and more—all adapting in real time based on the user's prompt.

How we built it

AutoForge’s architecture leverages a multi-agent framework with a central orchestrator that determines when new agents are needed and generates them on demand. Built primarily using llama index workflows, we have created agents for indexing personal folders and files from box.com and embedding them into pinecone, an example to-do list agent and we have extended some common functionality such as as searching the web, running custom code, etc using Toolhouse. We also used arize phoenix in the testing phase to see the diagnostics behind the scenes. The program's main feature stands for creating new agents and including them dynamically. We templatized the agent code and dynamically added the created Python file to the runtime to make the agent run.

Challenges we ran into

The main challenge was dynamically including new agents in the runtime. Given the limited time, we have created a templatized code to run agents and implemented an agent in the main workflow to dynamically run the generated Python code. Another notable challenge to the team was to learn and play around with llama-index and the set of APIs box.com and pinecone had to offer. Arize Pheonix did help us in this aspect.

Accomplishments that we're proud of

We’re proud to have successfully built a fully functional multi-agent assistant capable of dynamic task creation. AutoForge’s ability to generate specialized agents in response to complex requests was a major technical breakthrough, allowing us to demonstrate scalable, intelligent adaptability in action.

What we learned

Throughout this project, we learned the importance of agent orchestration and inter-agent communication. Building a system where agents collaborate rather than conflict was a rewarding challenge.

What's next for AutoForge

First, we host the project in the cloud to multiple users. Next, we plan to enhance AutoForge’s contextual understanding, allowing agents to adapt even more precisely to user preferences. We’re also exploring integrations with additional APIs to expand task types and plan to implement a more robust UI, making it easier for users to manage and visualize the agents working for them. This means building an agent evaluation phase later in the project for seamless dynamic agent integration in runtime.

Built With

• box
• llamaindex
• pinecone
• toolhouse