Agent experiment · 2025
Evolve Project
An agent that writes new Python tools and loads them while a task is still running.
Tool execution and reloading
- Task & context
- ModelChoose a tool
- ExecuteReturn its output
tools/new_tool.pyimportlib.reload()Available next turnExecution loop
- Conversation
- The model receives context, chooses a function, and uses its output in the next turn.
- Discovery
- The registry imports functions from
tools/, starting with shell execution and a log reader. - Backends
- The original loop uses Gemini. An experimental Ollama backend builds local-model tool schemas from Python docstrings.