🧬 CodeDNA: An In-Source Communication Protocol for AI Coding Agents
An in-source communication protocol where the writing agent encodes architectural context and the reading agent decodes it. The file is the channel. Every fragment carries the whole.
Compatible with:
CodeDNA is an inter-agent communication protocol implemented as in-source annotations. The writing agent embeds architectural context directly into source files; the reading agent decodes it at any point in the file. Like biological DNA — cut a hologram in half and you get two smaller complete images.
No RAG. No vector DB. No external rules files. Minimal drift (context co-located with code). 11 languages supported.
🎯 Less Prompt Engineering Needed: CodeDNA annotations help AI agents navigate the codebase with less manual guidance. Even less-technical users can get better multi-file fixes by describing the problem — the architectural context is already in the code.
The Problem
AI coding agents waste a significant fraction of their context window exploring irrelevant files, re-reading code, and missing cross-file constraints or reverse dependencies. The result: incomplete patches, higher token costs, and models that repeat the same mistakes across sessions.
The root cause is structural. Information like reverse dependencies and domain constraints cannot be inferred from a single file — they require reading the whole codebase. Without a way to persist that knowledge, every agent starts from scratch.
CodeDNA embeds this context directly in source files: used_by: maps reverse dependencies, rules: encodes domain constraints, and agent: / message: accumulate knowledge across sessions. It is not intended to replace retrieval systems, vector databases, or external memory — it provides a persistent architectural context layer inside the repository that any of those systems can build on.
This also enables agent-to-agent communication: a constraint discovered by Agent A is available to Agent B in a different session or a different model. Knowledge compounds in a versioned, inspectable form.
Preliminary results are encouraging:
- SWE-bench (single-agent navigation): +13pp F1 on Gemini 2.5 Flash (p=0.040), +9pp on DeepSeek Chat — zero-shot, no fine-tuning, just annotations.
- Multi-agent team experiments: 5-agent teams build complete applications 1.6x faster with CodeDNA;
used_by: contracts prevent integration gaps; the message: field (agent-to-agent chat) was adopted spontaneously in 98.2% of files.
- Fix quality (Claude Code): CodeDNA session matched 7/7 files of the official Django patch vs 6/7 for control, with zero failed edits.
Results are preliminary and require larger-scale validation.
Where CodeDNA sits in the AI memory stack