Agent
A model with a defined role, context and permitted tools—for example, proposing requirements from a brief.
AI ATELIER
A shared language for the work.
Practical definitions for architecture and interior design. Start with the task; use the technical term when it helps.
Agentic AI
AI that can select and execute permitted steps in a task, within controls set by the studio.
Orchestrator
The coordinator that assigns tasks, tracks state and stops at review gates. It does not inherit unlimited project access.
Multi-agent system
Several focused roles exchanging defined outputs, such as research, modelling and checking. More agents do not automatically mean better work.
MCP
Model Context Protocol: a standard interface for exposing tools and context to an AI application. Each connection still needs permissions and review controls.
API
A software interface for defined operations, such as retrieving object properties from a design application.
Context
The approved brief, constraints and relevant evidence supplied for the current task.
Project memory
Curated records that persist between tasks. Approved decisions stay separate from temporary agent notes.
RAG
Retrieval-augmented generation: finding relevant project records before asking a model to respond. Retrieved text is evidence, not authority to change permissions.
Tool calling
An agent requests a defined software action. The tool layer must enforce scope before executing it.
Human approval
An explicit decision by the responsible person, tied to a task or revision before consequential work proceeds.
Deterministic checking
Code applies a defined rule to valid model data, producing a reproducible result within stated assumptions.
Permissions
The enforced boundary of what a role can read, write or execute in a particular project.
Audit trail
A trace of sources, tool actions, model revisions and human decisions that helps reconstruct what happened.
Scheduler / cron job
Cron or an application scheduler starts the approved workflow; it does not decide the design. Configure the studio time zone, working days, concurrency lock, time limit and retry budget. A run records its source versions and completed steps so an interruption can be reconciled without duplicate writes. Missing sources, revoked access or failed checks produce an exception, not an empty success report. Client messages, model changes and calendar updates stay behind the agreed approval rules.
How to read contribution labels
These labels describe evidence attached to an output. They are not applied to an experiment merely because it is planned.
AI GENERATED
An AI system produced this output; the label alone does not imply review or approval.
AI ASSISTED
A person directed the work and used AI for identified parts of the process.
HUMAN DESIGNED
The design authorship and substantive design decisions are attributed to a person.
HUMAN APPROVED
A named reviewer approved a specific version for a stated purpose.
DETERMINISTICALLY CHECKED
Defined rules were executed and their scope, coverage and results are recorded. This is not blanket compliance.