AI coding agents are getting better fast, but I kept running into the same failure modes:
- they skip repository discovery
- they invent APIs, flags, or config keys
- they ignore existing project conventions
- they add unnecessary infrastructure
- they skip validation
- they treat unsupported checks as silently done
- they finish without a final codebase audit
So I built The Polyglot Protocol.
It is an open source senior-engineer workflow protocol for AI coding agents working across polyglot codebases.
Repo:
https://github.com/sabir-gbs/the-polyglot-protocol
What it includes
- guidance for 22 languages
- language-selection rules
- pre-codegen checklists
- do-not-generate policies
- validation scripts
- adapters for Codex, Claude Code, and OpenCode
- post-codegen audit scoring
The core idea
Before an agent changes code, it should:
- inspect the repository
- preserve existing conventions
- choose the right language and tooling
- avoid invented APIs or dependencies
- validate the result
- document unsupported checks as explicit
N/A - finish with a final audit
Model-agnostic by design
I have tested the protocol with frontier models as well as models from Qwen, Grok, Kimi, MiniMax, and others.
The point is not to depend on one model. Strong models still benefit from stricter engineering discipline, and smaller/local models benefit from clearer constraints.
Why open source it?
I want a practical, inspectable protocol that developers can adapt, critique, and improve.
If you use Codex, Claude Code, OpenCode, Cursor, Aider, or similar agents, I’d be interested in feedback:
- Is the protocol too strict?
- What guardrails are missing?
- What should be optional vs required?
- What would make this easier to adopt?