pycc

Development model · Public experiment

The project writes itself.

This pre-alpha project tests a strong form of AI-native software development: AI agents create the entire repository, while a human only manages the mission. The claim is about authorship, not the absence of human accountability. “AI-native” does not mean that pycc compiles AI models or an AI-specific language.

AI creates artifacts Human manages intent History public on GitHub

Authorship boundary

AI creates. The human manages.

AI agents Human manager
Write specifications and decisions Set goals and constraints
Implement every line of project code Choose priorities and product direction
Create tests, CI, and release automation Define success and acceptable risk
Review changes and maintain documentation Manage the experiment

The human does not handwrite patches into the project. No project code is handwritten by a human. The public Git history, pull requests, agent instructions, review artifacts, and session handoff logs make that process inspectable rather than asking readers to accept a hidden workflow. Git author and committer fields identify the accountable human maintainer, not the code generator; the AI-authorship claim is a maintainer attestation supported by those artifacts, not a fact provable from commit metadata alone.

Operating model

The repository is both product and memory.

  1. 1

    Specify. Agents read the project contract and turn the next goal into explicit, testable behavior.

  2. 2

    Prove red. A failing test or validation rule establishes the gap before the implementation claims success.

  3. 3

    Implement. Agents change code and documentation together, preserving the distinction between current and planned behavior.

  4. 4

    Review. Automated checks and an independent review loop look for regressions, unsupported claims, and unsafe process changes.

  5. 5

    Learn. Durable lessons become repository instructions or agent skills so later cycles start with more context.

Accountability

Autonomy still has gates.

The protected main branch requires an up-to-date pull request, required CI and audit checks, and resolved review conversations. Test coverage for the current Rust slice is enforced at 100% for both lines and regions. Documentation is part of the implementation contract, so a feature is incomplete while the specs describe old behavior.

Human-managed does not mean human-absent.

The manager remains responsible for goals, constraints, priorities, product choices, and the definition of acceptable outcomes. AI authorship is not a substitute for governance.

Audit trail

Inspect the experiment in public.

What the agents have shipped Current status → How the compiler is designed Architecture →