New 'Source Code Algebra' cuts LLM token use 10-100x for code changes
Ditch text editing—logical operations replace diff patches for AI agents.
Software engineering has long treated source code as plain text, but that mismatch creates friction—especially for LLM-based coding agents. Instead of materializing high-level plans as scattered text edits, Kevin Pulo proposes a radically different approach: source code algebra. Each operation corresponds to a single semantic change (e.g., renaming a function across files) and behaves like mathematical equation rewriting—supporting composition, nullipotency, and commutativity.
Pulo's proof-of-concept, SCAS, shows that LLM agents using these operations can complete non-local, cross-file changes with higher success rates and 10–100x fewer tokens compared to text-based baselines. The implication is clear: having LLMs emit algebraic operations rather than rewritten code is a promising direction for reducing token costs, improving reliability, and enabling more abstract reasoning in code generation. While preliminary, this work invites broader research into operator libraries, formal properties, and human-facing tooling.
- Proposes source code algebra: each operation makes a full semantic change (like equation rewriting) rather than text diffs.
- Proof-of-concept (SCAS) shows LLM agents achieve higher success rates with 10-100x fewer tokens for non-local cross-file edits.
- Operations exhibit composition, nullipotency, and commutativity—key properties absent in text editing that are ideal for agentic workflows.
Why It Matters
Slashing token use 10-100x while improving accuracy could revolutionize LLM-based coding agents and reduce costs dramatically.