AI ATELIER
Inside the AI Atelier.
A reference architecture for connected design workflows: what each component does, what it can access and where the designer decides.
The studio is the context.
Documents & drawingsHistorical projectsDesign journalsEmail & decisionsCalendar & deadlinesStandards & materials
One connected knowledge layer. Retrieval respects source permissions and returns the document, revision and date behind each answer.
The designer directs every stage.
Orchestrator / project managerPlans the work · assigns agents · tracks dependencies · routes approval · records the run
Select a step to explore its role.
01Client brief
The designer selects the current brief and approved references. Missing dimensions remain visible.
02Brief Agent
Proposes a requirement register with source links, questions and conflicts for review.
03Studio knowledge & project context
Retrieve the right information across the studio repository: current and historical projects, briefs, design journals, drawings, specifications, email and calendar. Results retain source links, dates and project permissions.
04Research / Spatial / MaterialsRSM
Parallel specialists explore different questions from the same context. An orchestrator manages their handovers.
05Human approval
The designer selects the direction and authorises the next modelling task.
06Modelling Agent
Prepares a candidate revision within the agreed dimensions and tool permissions.
07SketchUp / Revit / CAD / BIM
An assessed MCP, API or Python connection carries out bounded operations. Availability depends on the actual software interface.
08QA Agent + rule checks
Reports checked, failed and not-tested items. Code measures defined conditions; AI can explain findings.
09Human approval
A named reviewer checks evidence and approves a specific revision before release.
10Document / Visualise / Specify
Prepare reviewed drawings, draft schedules and views tied to the released revision.
Scheduled jobs / event triggers / manual runsAgreed time windows, scoped permissions, bounded retries and a review queue.
Studio workflow / human direction
Studio workflow / human direction
01Language models
Interpret project language and propose structured outputs. Select the model for the task and the data boundary.
Definition ↗ - What it can access
- Only the context explicitly supplied.
- What it can change
- Draft responses; no implicit write access.
- Where approval is required
- Before confidential inputs reach a new provider.
02Orchestrator / project manager
One coordinating role turns the approved brief into tasks, assigns specialists and maintains a visible record of progress, dependencies and decisions.
Definition ↗ - What it can access
- Task state and approved manifests.
- What it can change
- Task assignments, dependency state, run budget and review queue within the agreed scope.
- Where approval is required
- Before widening the scope or releasing a revision.
03Scheduled jobs and event triggers
Run a morning project briefing, refresh an approved document index or prepare an overnight model-check report. Each job has a trigger, an owner and a delivery point.
Definition ↗ - What it can access
- The approved job definition, source permissions, run state and working calendar.
- What it can change
- Starts scoped runs; records progress, exceptions and recovery. It cannot expand an agent’s permissions.
- Where approval is required
- Before enabling a new job, changing its scope or releasing a consequential write.
04Specialist agents
Give research, modelling and checking distinct inputs and acceptance criteria.
Definition ↗ - What it can access
- A task-specific context and tool subset.
- What it can change
- Assigned candidate outputs only.
- Where approval is required
- At design choices and inter-role handovers defined by the brief.
05Context and curated memory
A connected studio knowledge layer combines full repository search with the active project brief. Historical projects, journals, emails and calendar entries provide traceable context; approved decisions form curated memory.
Definition ↗ - What it can access
- Authorised repositories, project archives, selected mailboxes and calendars, filtered by user, project and source permissions.
- What it can change
- A separate search index and candidate memory records. Source documents, emails and calendar entries remain read-only unless an additional write task is explicitly approved.
- Where approval is required
- Before promoting an agent note to an accepted project decision.
06MCP / APIs / Python
Expose small, testable tool operations instead of uncontrolled application access.
Definition ↗ - What it can access
- Explicit endpoints and allowlisted files.
- What it can change
- Only permitted operations, checked before execution.
- Where approval is required
- Before enabling write operations or external communication.
07Design applications
Keep editable geometry in the studio’s actual authoring tools. Confirm version, API coverage and coordinate conventions during scoping.
Definition ↗ - What it can access
- A selected model revision and approved view set.
- What it can change
- Candidate models and exports in a separate workspace.
- Where approval is required
- Before a candidate replaces live project information.
08Validation and QA
Measure objective conditions with code; report failures and untested scope separately.
Definition ↗ - What it can access
- Actual model output and approved rule versions.
- What it can change
- Check reports; no silent model corrections.
- Where approval is required
- Before accepting exceptions or changing a rule.
09Permissions and audit trail
Limit each role’s authority and record who did what to which revision.
Definition ↗ - What it can access
- Role policy, operation metadata and approved logs.
- What it can change
- Controlled policy changes and append-only run records where implemented.
- Where approval is required
- Before granting access, changing policy or deleting records.
10Human approval
Place the designer inside the system as director, with explicit decisions and a recoverable release process.
Definition ↗ - What it can access
- Change sets, assumptions, check coverage and candidate outputs.
- What it can change
- Approve, reject or request a revision within professional responsibility.
- Where approval is required
- Required for consequential design decisions and release.
MCP: tools and resources ↗ Security and governance ↗