🚀 AutoGPT Roadmap - Empowering Agent Builders 👷

[Pwuts]

Objectives 🎯

• Let fellow builders benefit from our engineering efforts
• Enable fast and straightforward building/prototyping of AI-powered agents
• Minimize duplication of code, both for us and for other developers.

Background 💡

Turning research into practical applications

We want to enable straightforward building of (experimental) agents. Since the initial publication of AutoGPT, various agent and reasoning paradigms have emerged from experimentation and research. However, few have made it into PoCs with AutoGPT due to integration complexities. By improving the architecture, we can simplify experimentation (through modularity) and speed up building (through reusability of components).

Library code

AutoGPT contains a lot of library code - code that is not specific to our use case but essential for the system’s overall functionality. Whenever feasible, we try to import such code from third-party libraries. If suitable components are not available, it’s sensible to publish our own when we build a custom solution.

Vision 👉🏼

To make the above happen, each subproject in the AutoGPT project will have its own role:

• The generalist AutoGPT agent, with which we push the frontier of what’s possible with open-source agents
• Forge, an agent building SDK consisting of a component library, agent application templates, and tutorials on how to build your agent
• Our Frontend, which provides an easy to use interface for any agent that implements the Agent Protocol
• AGBenchmark, our tool for measuring agents’ capabilities

SWOT 🧭

Strengths 💪

• AutoGPT is a well-known project with a lot of reach

Weaknesses 🚧

• We don’t have a full overview of which components in AutoGPT could be replaced by external libs

Opportunities 🚀

• There is a gap in the market to be filled

Threats ⚠️

• There are already many LLM libs/frameworks
• Maintaining this could be a significant time sink

Related 👪

• 🚀 AutoGPT Roadmap - Vendor Liberty 🗝️ #6969

Actionables

• Separate Forge into component library and agent template(s)
• Test making an agent in Forge & update docs with findings
• Build example agent with Forge
• Move clear-cut library code from AutoGPT to Forge (or to /dev/null if Forge already has a better version)
  • Split forge.models.config into forge.models.config and forge.models.state #7000
  • Port autogpt.core.utils.json_schema from AutoGPT to Forge #7002
  • Port autogpt.core.resource.model_provider from AutoGPT to Forge #7001
  • Port autogpt.file_storage functionality from AutoGPT to Forge #7004
  • autogpt.agents.utils.exceptions
  • autogpt.core.prompting
  • autogpt.commands
  • A minimized version of autogpt.logs
• Reduce AutoGPT codebase size
  • Zealously replace own implementations by imported modules (third-party or from Forge) refactor(agent, forge): Move library code from autogpt to forge #7106
• Improve General Documentation for developers Auto-GPT developer documentation 📔 #5191
• Improve usefulness of abstractions
  • Add support for Image, Video, and Audio input into Forge and AutoGPT #7152