A working AI atelier.
Directed by your people.

Start with an input register: photographs, the design brief, supplied dimensions, available drawings and the intended output. Record which geometry is measured, which is supplied by the client and which is inferred. Missing scale or hidden conditions should remain a question or an explicit assumption.

For an early concept, some assumptions may be acceptable. The same assumptions may be unacceptable for detailed development. Define the intended use of each model stage before the agents begin. An attractive render does not resolve uncertainty in its underlying geometry.

A brief-and-reference agent structures requirements and open questions. A concept agent proposes spatial directions. The designer chooses a direction before a modelling agent develops it through approved software tools or Python routines. This keeps the proposal connected to a decision rather than letting the system produce detail indiscriminately.

The model should be editable and structured for the next task. Agree units, origins, naming, object hierarchy and the level of detail with the studio. A plausible surface is not necessarily a useful model. The modelling approach depends on the target software and available tool access, which are scoped before implementation.

A checking role reads the model state and evaluates defined rules. Deterministic code can test dimensions, object properties and other measurable constraints. AI can help interpret a brief or explain exceptions, but it should not substitute a fluent opinion for an executable geometric check.

The rendering role works from an identified model revision with agreed cameras, materials and output settings. Keep concept-image generation distinct from rendering the 3D model. They answer different questions, and a presentation should make the source of an image clear.

Give the checker the approved requirements and the model output, rather than only the modeller’s description of what it did. Return a list of detected exceptions, unsupported assumptions and checks that could not run. A pass means the defined tests passed; it is not proof that every aspect of the design is correct.

The coordinator can send a bounded correction task back to the modelling role. Limit iterations and escalate unresolved conflicts. Require designer approval where a correction changes design intent or where a missing dimension would otherwise be guessed.

Maintain versions and checkpoints so the team can compare alternatives and return to an approved state. Restrict model-writing tools to the intended project and keep destructive operations behind explicit authorisation. Render from a known revision, not whichever file happened to be saved last.

The meaningful gates are input acceptance, concept selection, model development and presentation release. Between them, specialist agents can prepare and execute authorised tasks. Project leads remain responsible for interpretation and sign-off.

Shared context carries the approved brief and decisions; curated memory preserves the rationale. Python tools perform modelling and checks. Cron jobs can prepare queues or run agreed overnight tasks, while logs and permissions make the activity traceable. The orchestration connects these capabilities without giving every agent unrestricted access.

We scope a first workflow around one model type and a defined deliverable, then test fidelity, editability, checking quality and review effort. Expansion follows evidence. This is how a multi-agent atelier becomes a useful studio capability, rather than just a convincing demonstration.