The six layers
People, Data, Agents Tools and Skills, Models, Security, Integration. The spine of the platform and of these docs.
Adam AI is built as six layers. The split is not decoration: every control, every integration and every feature lives in exactly one layer, and the boundaries between layers are where the security and the no-lock-in promises are enforced.
The stack at a glance
| Layer | Question it answers | Owns | Never does |
|---|---|---|---|
| 1. People | Who may see and do what? | Users, roles, Markings, approvals, the human checkpoint | Hold credentials; act without a named owner |
| 2. Data | Where is the data and how is it classified? | Company knowledge, records, documents, Markings on data | Leave the customer's estate; get copied into a model |
| 3. Agents, tools and skills | What work gets done, and how? | Sandboxed agents, 1,000+ skills, the Tools arsenal, My Builds | Hold standing access; run outside a session |
| 4. Models | Which intelligence answers? | Local and hosted models, the model gateway, routing and evals | See data it was not handed; retain anything |
| 5. Security | How is every action gated and recorded? | Session-based least privilege, the single checkpoint, audit | Trust an agent's own description of itself |
| 6. Integration | How does Adam AI meet the systems you already run? | Connectors, MCP servers, the fleet, the catalogue | Sync a customer's local integrations; harvest secrets |
Why layers
The single biggest risk in agentic AI is not a bad answer. It is an agent spawned with the highest level of access its creator happened to have, talking to thirty systems through six different tools. An average enterprise runs 30+ systems. Six AI tools against them is 180 ways in, most of them delegated to agents nobody is watching.
Layering means every one of those paths has to pass through the same place. Agents (layer 3) get access per session from the Security layer (5), scoped to the Markings (1) on the data (2) they are allowed to touch, reaching real systems only through the Integration layer (6). The model (4) only ever sees what the agent was handed. One checkpoint, not 180.
How the layers talk to each other
- A person in the People layer states an outcome and holds the Markings that bound what may be touched.
- The Agents layer proposes a build. It cannot start without a named owner and a proposal.
- The Security layer issues a session: the smallest set of rights needed for this task, for this long, recorded.
- The Data layer answers queries filtered by the session's Markings. Anything derived inherits the Markings of its sources.
- The Model layer receives only the filtered context and returns a result. Nothing is retained.
- The Integration layer carries actions out to the real systems and brings results back. Every call is attributed to the session.
Architect sits across all six
Architect is the module that keeps each layer swappable. Change the model, change the tools, change the products. Even change us. The only lock-in is the growth.
Where to go next
- Owners and managers: People then Security.
- Technical buyers: Models then Integration.
- Engineers: the Engineer reference, starting with Stack and deploy.