The AI-Native Framework
An AI-native operating system for building and running product-led companies. spec-driven · human-governed · provider-agnostic
The question
What happens when a company runs every function — not just engineering — as governed agentic work?
What it is
Most “AI for X” today is bolt-on automation: a chatbot here, a copilot there, a draft generator somewhere else. This project takes a different bet — it treats AI as the substrate the company runs on, and codifies the operating loop end-to-end: intent → discovery → product → build & ship → go-to-market → operations → feedback & learning.
The framework is designed so agents can execute structured work under explicit constraints, while humans retain authority for strategy, ambiguity, and high-stakes decisions. The target is high leverage, not fake autonomy — a 90/10 automation-to-human ratio that only holds when judgment checkpoints, confidence thresholds, and escalation rules are explicit.
The repository is both the artifact and the experiment: spec-driven artifacts, event-observable runtime, provider-agnostic interfaces, and human governance baked into the authority ladder rather than left to chat.
How it works
Four layers an agent or a person can read top-to-bottom.
- 1
Specs are the source of truth.
A product is described as a YAML spec, validated by a JSON schema. No feature exists until it exists in the spec. The spec declares slices, events, and policies.
- 2
Interfaces describe capabilities, not vendors.
Logical operations — read_spec, validate_spec, emit_event — are defined as contracts. Anything that talks to the framework (a model, an IDE, an MCP adapter, a CI job) goes through them.
- 3
The agent surface tells agents what to do.
A root AGENTS.md routes through three indexes: SKILLS (the roles an agent can take), PLAYBOOKS (procedures for recurring work), and MEMORY (durable repo facts and open loops).
- 4
Playbooks turn repeated work into reusable procedures.
Each playbook is short, self-contained, and executable by an agent without further context. Feature implementation, PR execution loop, release, incident response — all live as governed procedures.
Built with
Status
Active exploration. Independent project. Open-source.