Agent Frameworks

AI
gpt
framework

Agents are semi-autonomous programs that can adjust their behavior based on what they encounter.

Agents are more difficult to build than they look, partly because they suffer from a single-point-of-failure problem. Aurorean explains how these systems success rate declines precipitously the more passes they make:

Even in scenarios where a system is 99.99% reliabile, it still leads to 10,000 errors over the course of 1 billion interactions, and this is an unacceptable failure rate for manufacturing processes, public health settings and other critical domains in society.


Zhou et al. (2023) says this is hard to get right, and that current systems perform poorly.

We experiment with several baseline agents, integrating recent techniques such as reasoning before acting. The results demonstrate that solving complex tasks is challenging: our best GPT-4-based agent only achieves an end-to-end task success rate of 14.41%, significantly lower than the human performance of 78.24%.

Commercial Applications

CrewAI is “the complete platform”: Build and deploy automated workflows using any LLM and cloud platform.

General Discussion

Zvi (Don’t Worry about the Vase) offers a summary on AutoGPT:

Zvi’s summary of how AutoGPT works

Lilian Weng has an in-depth description at LLM Powered Autonomous Agents

Several proof-of-concepts demos, such as AutoGPT, GPT-Engineer and BabAGI, serve as inspiring examples. The potentiality of LLM extends beyond generating well-written copies, stories, essays and programs; it can be framed as a powerful general problem solver.

Microsoft AutoGen

AutoGen is a GitHub-based framework for development of LLM applications using multiple agents that can converse with each other to solve tasks. AutoGen agents are customizable, conversable, and seamlessly allow human participation. They can operate in various modes that employ combinations of LLMs, human inputs, and tools.

AutoGen offers the following key features:

  • Asynchronous Messaging: Agents communicate via asynchronous messages, supporting both event-driven and request/response interaction patterns.
  • Full type support: use types in all interfaces and enforced type check on build, with a focus on quality and cohesiveness
  • Scalable & Distributed: Design complex, distributed agent networks that can operate across organizational boundaries.
  • Modular & Extensible: Customize your system with pluggable components: custom agents, tools, memory, and models.
  • Cross-Language Support: Interoperate agents across different programming languages. Currently supports Python and .NET, with more languages coming soon.
  • Observability & Debugging: Built-in features and tools for tracking, tracing, and debugging agent interactions and workflows, including support for industry standard observability with OpenTelemetry

AutoGPT

AutoGPT is a framework for creating a GPT-powered agent. You define a series of scripts that interact with the OpenAI API to generate text, save or process the results as necessary, and then send followup queries.

AutoGPT.net is a whole web site devoted to using AutoGPT

Installation and how to get started

Twitter thread by @hasantoxr

AutoChain

AutoChain

AutoChain takes inspiration from LangChain and AutoGPT and aims to solve both problems by providing a lightweight and extensible framework for developers to build their own agents using LLMs with custom tools and automatically evaluating different user scenarios with simulated conversations. Experienced user of LangChain would find AutoChain is easy to navigate since they share similar but simpler concepts

MetaGPT

MetaGPT: The Multi-Agent Framework > MetaGPT: Simulates a whole software company 🤔 > > It assigns roles like product managers, architects, project managers, and engineers to GPTs. > > With just one line of code, MetaGPT generates user stories, competitive analyses, requirements, data structures, APIs, documents, and more.

Software Company Multi-Role Schematic (Gradually Implementing)

via TuringPost

GPT Researcher

GPT Researcher

The agent can produce detailed, factual and unbiased research reports, with customization options for focusing on relevant resources, outlines, and lessons. Inspired by AutoGPT and the recent Plan-and-Solve paper, GPT Researcher addresses issues of speed and determinism, offering a more stable performance and increased speed through parallelized agent work, as opposed to synchronous operations.

MiniChain

MiniChain

This chaining tool might be the first actually truly useful one for both prototyping and production. It uses function decorators and YAML templates in a clever and powerful way to enable chaining. Examples show that you can write a chat bot, vector database, and more in just 20 lines.

Langchain

Langchain, developed by Harrison Chase, is a Python and JavaScript library for interfacing with OpenAI’s GPT APIs (later expanding to more models) for AI text generation. More specifically, it’s an implementation of the paper ReAct: Synergizing Reasoning and Acting in Language Models published October 2022, colloquially known as the ReAct paper, which demonstrates a prompting technique to allow the model to “reason” (with a chain-of-thoughts) and “act” (by being able to use a tool from a predefined set of tools, such as being able to search the internet). This combination is shown to drastically improve output text quality and give large language models the ability to correctly solve problems.

and Langchain in Realworld, a detailed list of use cases and how to build them.

LangChain Templates

and LangGraph : a library for building stateful, multi-actor applications with LLMs (Highly recommended by AI Tinkerer Jay Bartot - “Otto - A Virtual Travel Agency”. https://linkedin.com/in/jay-bartot)


Max Woolf’s explanation of why LangChain isn’t necessary: “LangChain’s vaunted prompt engineering is just f-strings, a feature present in every modern Python installation, but with extra steps.”

(and see HN)


Brighten.ai

See also

llama-agents is an async-first framework for building, iterating, and productionizing multi-agent systems, including multi-agent communication, distributed tool execution, human-in-the-loop, and more.


Artificial intelligence is algorithmic mimicry: why artificial “agents” are not (and won’t be) proper agents Johannes Jaeger 2024

References

Zhou, Shuyan, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, et al. 2023. WebArena: A Realistic Web Environment for Building Autonomous Agents.” arXiv. http://arxiv.org/abs/2307.13854.