Souveraineté sur son travail avec les agents
The state in which your work with agents keeps existing if the provider, the model or the tool disappears. The word covers three distinct independences that often get conflated: 1. **Provider independence**: your work does not die if Anthropic or OpenAI changes its terms. 2. **Model independence**: it does not assume one specific model to be understood and resumed. 3. **Environment independence**: it does not live inside an application conversation history. Claude Code, the Claude app, ChatGPT, Codex are environments, not storage. Sovereignty is not declared, it is observed in three places: - **Architecture**: the working logic that survives between two sessions (memory, orchestration, declared skills). - **Trace**: the shape and filing of your deliverables. Where they live, in what format, whether they are versioned. - **Doctrine**: your expertise put in a form an agent can read and navigate, not just you. If one of the three is missing, what you own is not a system yet.
Strengths
- Verified by a concrete manipulation, not by a statement of intent
- Compatible with proprietary models: it is the filing of the work that decides, not the hosting
- The three-way split (architecture, trace, doctrine) says where to start when everything is missing
Limitations
- The word has become a marketing argument: many use it to sell local hosting, which is only one of the cases
- Chased too early, it builds infrastructure around work that has no value yet
Best for
- Deciding whether an agent system is ready to survive a provider or model change
- Arbitrating what gets written into files rather than into a conversation