Build a simple learning agent

In this project, we will build a simple learning agent that

• Given an outcome description (e.g. my-project.outcome.md), the agent will check a user's work against the outcome and
  • Create a learning path for the user to follow to achieve the outcome (e.g. my-project.path.md)
  • Track the user's progress at an interval (e.g. every 30s) to see what's changed and update the progress
  • Give the user feedback on demand (e.g. "How am I doing?" what should I do next?)

Setup

• Create the outcome file: terminal.guide.outcome.md
• Create the working directory: terminal-guide/
• Create your agent script: learning_agent.py
  • setup API keys in the .env file
  • setup the venv folder
  • install the dependencies
  • make your script executable: chmod +x learning_agent.py
  • checkpoint functionality
    • create a simple cli command e.g. learning_agent.py checkpoint that will look at the outcome, compare with work so far, give next steps and feedback

Terminology

• Agent: A program that can act on its own behalf.
• Context: The information that the agent uses to act.
• Harness: The program that runs the agent and provides it with the context it needs to act.
• Loop: The program that runs the agent and provides it with the context it needs to act.
• Turn: The sequence of steps between an initial user prompt inference and final agent output/response.
• Tool: A function that the agent can call to perform a task.

Reading

• Unrolling the Codex Agent Loop