# User-Controlled AI Workspaces: Persistent Projects + Multi-Agent Workflows
I would like to suggest a broader evolution of ChatGPT: a **user-controlled AI workspace** where multiple specialized agents can work together on long-running projects, using the user’s own files, knowledge bases and external services.
The key principle would be:
> **The user owns the workspace and the data. AI agents operate within it under explicit permissions.**
## Why this would be valuable
Today, ChatGPT is extremely useful for individual tasks, but long-running professional projects still require users to manually provide context, locate documents, manage versions and coordinate different tasks.
A persistent project workspace could change this.
For example, an architecture project could contain:
* DWG/CAD files
* BIM models
* PDFs
* photographs
* surveys
* legislation
* emails
* notes
* previous versions
* decisions
* pending questions
The AI would not need to permanently store all of this inside ChatGPT. It could work with data stored in a user-controlled environment such as a NAS, local server, Google Drive, OneDrive, SharePoint or another MCP-compatible service.
## A coordinator agent
The user would interact primarily with a **Project Coordinator Agent**.
The coordinator would understand the user’s objective and decide which specialized agents or tools should be involved.
For example:
**Architecture Agent**
→ develops spatial solutions and analyses plans.
**Regulatory Agent**
→ researches legislation and checks compliance.
**GIS Agent**
→ analyses location and territorial constraints.
**Document Agent**
→ searches project documentation and tracks relevant information.
**BIM Agent**
→ works with BIM data and models.
**Communication Agent**
→ prepares emails, reports and other documents.
The user would not need to manually select the appropriate agent.
## Shared project memory
All agents should work from a common project context containing:
* confirmed facts
* decisions
* assumptions
* unresolved questions
* sources
* previous outputs
* approvals
* project history
This would allow the system to maintain continuity throughout a project rather than treating every conversation as an isolated interaction.
## User-controlled permissions
This would be essential.
Agents should have granular permissions such as:
**Read**
→ inspect files but cannot modify them.
**Create**
→ create working documents.
**Modify**
→ modify files within an authorised workspace.
**External action**
→ send emails, update systems or perform other actions only when authorised.
For example:
`/Project/Official/`
could be read-only, while:
`/Project/AI_Working/`
could be writable by agents.
Changes to official project files could require explicit human approval.
## Human approval checkpoints
The system should distinguish between tasks that can be performed autonomously and decisions requiring human approval.
AI could:
* research
* analyse
* calculate
* compare
* create drafts
* generate alternatives
* perform checks
Human approval could be required before:
* issuing official documents
* sending client communications
* modifying approved project files
* submitting documents to authorities
* making financial commitments
## Verification between agents
The system could also use agents to verify each other’s work.
For example:
**Architecture Agent**
→ proposes a floor plan
↓
**Regulatory Agent**
→ checks applicable requirements
↓
**Geometry Agent**
→ checks dimensions and areas
↓
**Critical Review Agent**
→ searches for inconsistencies
↓
**Coordinator**
→ presents the result together with sources, assumptions and unresolved issues
This would make multi-agent workflows considerably more reliable.
## The bigger idea
The goal is not necessarily to create one enormous AI that knows everything.
It is to create a persistent environment where:
**User + Projects + Data + Memory + Tools + Specialized Agents + Verification**
work together.
The underlying AI models could evolve over time without forcing users to rebuild their entire digital workspace.
The user’s data, project history and permissions would remain under the user’s control.
### The principle
> **Don’t make the AI own the user’s knowledge. Make the AI capable of working intelligently with the user’s knowledge.**
I believe this could be an important step in moving ChatGPT from a conversational assistant toward a genuine long-term professional collaborator.