AI Alpha Squad started as a question: what if a small team of specialized agents could run a full delivery cycle — from a vague business ask to a tagged release — with a human Director keeping final authority?
This is an open experiment, not a finished product. It is a governed pipeline where every decision is auditable on GitHub Issues and every agent has a narrow mandate. Business Owner analyzes and seeks approval. Architect designs. Developer ships code to a target repository. QA, Security, DevOps, and Tech Writer validate in parallel. Release Manager gates the release.
Infrastructure & models
The squad runs on GitHub Actions as the runner and Hugging Face Inference Providers for the models behind every agent. Planning roles — Business Owner, Architect, QA, Security, Tech Writer, Release Manager — call the
Hugging Face router and post their deliverables straight to the issue. Developer and DevOps run an Actions agent that clones the target repo and drives a Hugging Face tool loop to open a PR. Models are configurable per agent via SQUAD_HF_DEFAULT_MODEL and per-role overrides, with cost-first provider routing by default.
A GitHub Copilot coding-agent path still exists as a legacy, optional runtime, but Hugging Face on Actions is the default and the direction we build toward.
The lab is intentionally experimental. We iterate on orchestration, agent prompts, open models, and runtimes while keeping the same contract: Issues are truth. Releases need Director sign-off. WhatsApp is for speed, not for secrets.
Delivery flow
From the Director's first signal to a production release on the target repository: