An agent in 100 lines of Lisp

18 hours ago

The author recalls a university AI course around 2000 focused on Lisp as the language for symbolic AI, which lacked modern elements like neural networks or deep learning.
Reflecting on Lisp's elegance for recursion, the author connects it to building AI agents, proposing that an agent loop is essentially a recursive function.
A simple 8-line Common Lisp function demonstrates an agent loop where messages are processed, tools are executed if needed, and recursion handles the state.
By leveraging Lisp's homoiconic nature (code as data), the agent uses a single eval tool to run arbitrary Lisp code, allowing dynamic function creation and execution.
Memory persistence is achieved in 20 lines by serializing message lists to/from a file, enabling conversations to survive across sessions without complex dependencies.
The agent dynamically built its own web search function using eval, showcasing an open tool catalog where capabilities are created at runtime rather than pre-defined.
Skills are stored as memories in transcripts, contrasting with traditional agent platforms that fix tools at design time, suggesting a more flexible, human-like learning approach.
The author concludes that while symbolic AI lost to deep learning, Lisp's strength in manipulating programs as data aligns well with modern AI agent engineering.

Hasty Briefsbeta