Ask a team how they organise their work and nobody says “it’s all one big document.” They describe layers. The company has a brand voice and a compliance posture. The team has its rituals, its vocabulary, its way of writing. The client has a history, a red-lines list, an account lead. The project has a brief, a decisions log, open questions. The individual has a role, a writing style, preferences.
Those five layers are how a team already thinks. Sempleo models context on the same five layers, because the shape of the model should match the shape of the work.
We built this on purpose. Here is why each layer matters, and why collapsing them into one pile — as most “AI context” products do — loses the thing that makes context useful.
Company is what every agent should know before saying a word. Mission, brand voice, escalation policy, legal posture, product descriptions. This is the layer Brand and Legal own, and it should be locked for everyone else. An agent that writes customer-facing copy without reading the company layer is producing generic work. An agent that reads it — and respects the signoffs, the forbidden phrases, the escalation triggers — is producing your firm’s work.
Team is how your team operates. Your sprint cadence. Your definition of “done.” Your vocabulary — what you mean when you say “a brief,” which is different from what another team means. Your rituals. This layer is owned by the team lead, proposed by anyone on the team, reviewed on its own cadence. Agents map generic terms to these team-specific ones before writing anything.
Client is everything known about an external counterparty. History, preferences, contracts, stakeholder map, voice samples. This is the layer the account lead owns. It carries red lines. It carries a tone. An email drafter that writes to the counterparty without the client layer will sound like a template. With it, it will sound like you.
Project is the story-so-far behind a body of work. The brief. The decisions log. Open questions. Milestones. A status writer that reads the project layer can produce the update your PM would have written. A status writer that does not will produce a summary of Jira tickets.
User is per-person. Role, voice, signoff style, preferred formatting. The user layer is what makes the same agent output sound different depending on who asked. It is small but it is load-bearing — without it, every piece of work has the same writer.
Each layer has its own owner, its own review cadence, its own authority level. Entries inside a layer carry a quality dot — empty, thin, or sharp — scored by Haiku, and a freshness stamp set per field by the layer owner. When an agent runs, it resolves the layers it needs, reads the fields, and cites them in the output. When a field is thin, the agent surfaces the gap rather than guessing. When a field is stale past its freshness window, same.
This matters because flattening context is a local optimum with a long tail of failures. A single pile of documents can power a demo. It cannot maintain the separation of concerns a real organisation runs on. Brand does not own the client record. The account lead does not own the company voice. Everyone reviews on a different cadence. An ontology that pretends otherwise produces output that is plausible and wrong — the worst failure mode an enterprise AI product can have.
Layers are how teams already work. Sempleo’s ontology matches. That is the whole design.
If you want to see what this looks like in practice, the context page walks through each layer with a designed example. If you want to sit with me and shape how it works for your team, the founding-customer cohort is open.
