Security
Control is not an extra feature. It is the architecture.
The intelligence of the model is useful precisely because software bounds it. Authorization, money and critical rules never depend on what the model understood.
The agent does not have unlimited authority.
Intelligence is probabilistic; authority is deterministic. The model interprets, the code grants permission.
least privilege
The agent gets only the access the task requires, and nothing beyond it.
deterministic rules
Critical rules, limits and authorization live in code, not in a prompt.
human approval
Actions with financial, commercial or operational impact go through human approval.
audit trail
Every relevant execution is recorded, with who, when and what came out of it.
cost limits
Token budgets and a per-session cost ceiling are part of the architecture.
observability
A single trace connects the request, the model, persistence and the final effect.
Practices
What that means in practice.
- Data
- We collect the minimum required. Secrets and keys stay server-side, never in the browser or the prompt. Telemetry carries identifiers, not personal content.
- Agents
- The agent does not run SQL, does not browse freely, receives no arbitrary tools, and never decides when a rule may be broken.
- Access
- Administrative access goes through corporate identity with strong authentication, and every meaningful change writes an audit record.
- Retention
- Retention periods are explicit configuration, reviewed legally, not an accidental consequence of the database.