Manifesto
“A project without process is a conversation. Not a delivery.”
The observation
Section titled “The observation”A large part of the AI industry copies human organization — agents with personas, roles, titles. This approach has the merit of being legible. Our experience suggests it isn’t always the most robust mechanism when you’re after quality, memory, and consistency over time.
An AI agent has no fatigue, no ego, no office politics. But its context window is limited, and it generally forgets everything between sessions.
A shift
Section titled “A shift”Rather than managing agents, you can also define the framework in which they operate.
Agents are like Kubernetes pods — stateless, scalable, without identity. You don’t name a pod. You define the desired state and the orchestrator allocates resources.
The 7 founding principles
Section titled “The 7 founding principles”- The manifest comes first. Everything starts from what the human defined.
- Agents are formless. No required persona, no fixed role.
- Memory is alive. It grows with each sprint — and it consolidates.
- Control is human. Agents execute. The human validates and steers.
- Complement, don’t mimic. AI is not a virtual human.
- Progressive by design. Start small, add when needed.
- Sovereignty through text. Project knowledge lives in files you own, version, review, and migrate — not inside a vendor.
The human’s role
Section titled “The human’s role”The human is the system architect. They define the foundation: the skills, the rules, the structure, the criteria. This work demands an intimate understanding of the project, its stakes, its business context — an understanding that remains human.
The human doesn’t necessarily write every line of code. They define the frame in which the team and the AI produce together.
Why text?
Section titled “Why text?”The manifest is markdown. Skills are markdown. Memory is markdown. Rules are markdown.
The AI is an interchangeable engine. The foundation is what persists.
This does not guarantee identical outputs across models. It guarantees continuity of context, process, and memory when the engine changes.
The day a vendor changes its terms, raises its prices, or ships a model that breaks your workflow — you switch the engine. You do not start over.