Agent Meta-Programming System Runtime

Amps is a research project to develop a language and runtime for self-improving AI agents. Unlike many existing agent architectures, Amps agents operate in a reflective programming environment which allows them to systematically understand and modify their own behavior systems at any level of abstraction.

This project is led by Daniel Sosebee and builds on the concept of human programming as described and developed in Methodable. For questions or ideas, please reach out via email.

Agenda

1. Initial research - review existing literature on AI agents, planning systems, and human executive functioning.
2. Problem identification - collect problems that will serve as a realistic test of the efficacy of autonomous AI agents.
3. Human trial recording - record a human (mysel) solving the problems identified in step 2 out loud, on video and in screen recording.
4. Pseudocode development - Write pseudocode programs based on human trial data to symbolically describe the human perceptions, thoughts, and actions.
5. Language spec development - Inspired by the human pseudocode, write a simple language spec for the Amps language.
6. Initial Python library development - develop a Python library that implements the Amps language.
7. Iterative system improvement - execute test runs against the problem list and record success metrics. Use the results to iteratively improve the system and reach a v1 language spec.
8. Fine-tuning data collection - collect a dataset of Amps contexts and correct outputs.
9. LLM fine-tuning - fine-tune an open-source language model on the dataset in order to better generate Amps commands.
10. Full implementation and evaluation - implement the full Amps runtime and evaluate it on the problem list.
11. Documentation - write up the results and publish online.

Related materials

The following are not associated with Amps but will serve as inspiration and reference for the project.

Papers

• 2024-03-23: Shared Protentions in Multi-Agent Active Inference
• 2024-03-14: Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking
• 2024-01-08: Mixtral of Experts
• 2023: Artificial Intelligence: Foundations of Computations Agents
• 2023-06-28: Curious Replay for Model-based Adaptation
• 2023-06-20: RoboCat: A self-improving robotic agent
• 2023-05-24: Voyager: An Open-Ended Embodied Agent with Large Language Models
• 2023-05-19: CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing
• 2023-05-17: Tree of Thoughts: Deliberate Problem Solving with Large Language Models
• 2023-23-20: Reflexion: Language Agents with Verbal Reinforcement Learning
• 2023-01-25: Reflective Artificial Intelligence
• 2022-12-15: Transformers Learn In-Context by Gradient Descent

See also Spinning Up in Deep RL which cites many classic RL papers.

Open source projects

• babyagi
• AutoGPT
• AgentGPT
• langgraph

Companies

Many general intelligence companies are approaching the problem from a software engineering perspective.

• Cognition is building "the first AI software engineer".
• Imbue is "creating practical AI agents that can accomplish larger goals and safely work for us in the real world."
• Magic is "working on frontier-scale code models to build a coworker, not just a copilot."
• Altera is "partnering with game studios and designers to power agent first games."