🚀 AutoGPT Roadmap - Ability Acquisition 💡

[Pwuts]

Objective 🎯

AutoGPT should be able to adapt to requirements / user requests that are outside the (practical) scope of its built-in/existing abilities.

SWOT 🧭

Strengths 💪

• Dynamic planning with LLM + execute_python_code allows solving problems with one-off code execution

Weaknesses 🚧

• execute_python_code is limited
  • it does not persist variables
  • its execution environment has a limited set of libraries available
• No procedural memory → recurring necessity for custom code execution requires repeated use of execute_python_code

Opportunities 🚀

• Procedural memory (e.g. as demonstrated by Voyager; FunSearch)
• Allowing the agent to write code with LLM calls could significantly increase the agent’s ability to adapt

Threats ⚠️

• Ability acquisition could be the kryptonite that enables AutoGPT to do something scary

Related 👪

• 🚀 AutoGPT Roadmap - Workflow efficacy 🧠 #6964
  • implement → execute → evaluate

Actionables

• Implement basic Procedural Memory (including procedure generation)
  • Generating procedures/code based on intent
  • Looking up + retrieving procedures based on intent
  • Executing retrieved procedures
  • Opt-in global shared Procedural Memory
• Add ability to generate & use API client from OpenAPI spec
• Integrate with Jupyter(Hub) for a more effective code execution environment
  • Set up for non-trusted users
• Add LLM access to code execution environment
  • Decide: Provide simplified interface OR direct access to LLM client?
  • Implement access to LLM in code execution environment
  • Instruct the agent on the availability of LLM services