TS-Agents 🤖

TS-Agents is an agentic framework for TypeScript. It allows you to build multi-agent systems with ease. It allows creating agents, providing them with tools and memory and creating flows to orchestrate them. Graph based flows allow you to create complex multi-step / decision trees are the next step.

Why?

I became interested in creating custom agent flows and running them locally with recent advancements in LLMs and new models, like DeepSeek that seemed lightweight enough to run on my MacBook.

While there are multiple Python 🐍 based frameworks, JS and TS specifically were less mature. I took that as an opportunity to learn more about the field, build something of my own, and deepen my understanding by getting my hands dirty.

By no means I consider this project production ready, it's more of a proof of concept and a playground for me to learn more about the field. However, I'll be happy to receive feedback and contributions and improve the project.

How Does It Work? ⚙️

The framework is built around the concept of agents. Agents are the main building blocks of the framework. They are responsible for performing tasks and using tools.

Building a flow is done by passing a list of agents to be executed. Each agent is passed an LLM instance and a list of tools. They can also have previous agents passed as context. Flow can utilize different LLMs to execute different agents.

LLM class is using a connector to connect to the LLM provider. Currently OpenAI and Ollama 🦙 are supported, but you can also pass a custom LLM connector.

Examples:

• OpenAI - Simple example of using OpenAI
• Parallel - Example of executing multiple agents in parallel
• Multiagent - Example of using multiple agents that feed into each other in a flow

All of the examples are using one LLM instance for all agents. You can tweak it by changing model parameter in the LLM class constructor located in each agents.ts file.

Project Goals 🚀

• ✅ Create a basic framework to connect with LLMs and execute simple tasks
• ✅ Create an agent that can be instructed to perform a task using provided LLM as a base
• ✅ Give the agent access to tools and performing tool calls
• ✅ Add example tools
• ✅ Create a flow where multiple agents can be executed in series
• ✅ Add the ability to use previous agents' repsonses as context for the next one
• ✅ Add an option to execute multiple agents in parallel
• 🛠️ Crate a graph based flow to orchestrate complex agent flows
• 🛠️ Add a memory that agents can reach into for context in graph flows
• 🛠️ Add a fine-tuning ability

Next Steps 🚧

• Add graph based flows (start, end, task)
• Add decision making nodes
• Add map-reduce nodes
• Add memory for graph flows
• Add a way to pass fine-tuning data to LLM models
• Add a way to generate training data for LLM models
• Add more tooling
• Add more examples
• Add more LLM connectors

How to Get Started

If you want to use OpenAI, copy the .env.example file to .env and add your API key. If you want to use Ollama, make sure it is running and the model(s) you want to use is downloaded.

Install dependencies:

yarn install

Run the example:

yarn example:openai
yarn example:parallel
yarn example:multiagent

Structure

ts-agents
├── lib
│   ├── connectors/      - Connectors for the LLM providers
│   ├── agent.ts         - Agent class
│   ├── flow.ts          - Flow class (orchestrator)
│   ├── llm.ts           - LLM class
│   ├── tool.ts          - Tool class
│   └── types.ts         - Types for the framework
├── models/              - Custom models for Ollama
├── tools/               - Custom tools for the agents
└── example              - Example flows
    ├── openai/
    ├── parallel/
    ├── multiagent/
    └── metadata/

Contributions

TS-Agents is open for contributions and feedback! By no means am I the expert in the field and I'll be happy to take suggestions. If you have ideas, improvements, or bug reports, feel free to create an issue or open a pull request.