Mustafa Qureshi.ai
In active development AI Lab

An Intelligent Workforce.
One Personal Command Center.

Alef AI is an evolving multi-agent AI system designed to coordinate specialised digital agents, understand tasks, manage workflows and assist with complex personal and business operations.

Not a chatbot that answers one prompt at a time.

The long-term vision is a digital workforce: agents that each own a job, hand work to each other, and report back — helping with research, planning, software development, business operations and desktop workflows.

It is being built to be capable but controlled. Consequential actions wait for approval, and Alef AI is not given unrestricted access to accounts, money, private data or devices.

01Agent Town

A virtual office for digital agents.

Watch a simulated task move through the team — assignments, handoffs, a meeting, a review and an approval request. Click any agent to see what it's doing. When approval is needed, you decide.

alef://agent-townDemonstration · simulated
Command desk
Meeting room
Shared workspace
Owner ↑
Step 0 / 9
02Multi-agent architecture

One coordinator. Seven specialists.

Each agent has defined responsibilities, permissions, policies and performance tracking. Not every agent runs fully autonomously — this is a system being built and tested.

Boss Agent

The central intelligence. It owns the request from start to finish.

  1. 01Understand the request
  2. 02Break down the task
  3. 03Assign specialist agents
  4. 04Manage dependencies
  5. 05Coordinate execution
  6. 06Collect results
  7. 07Review outputs
  8. 08Report progress
  9. 09Request approval for consequential actions
  • Planner AgentBreaks complex requests into structured plans.
  • Researcher AgentGathers and analyses information.
  • Operator AgentHandles approved computer and workflow operations.
  • Curator AgentOrganises information and knowledge.
  • Customer Support AgentSupports customer-service workflows.
  • Finance AgentAssists with finance-related business workflows.
  • Administration AgentAssists with administrative operations.
See the full architecture
03Capabilities

What's built, and what's still being integrated.

Foundation builtIn developmentFuture direction
01

Desktop intelligence

A desktop automation foundation with a Windows-oriented interface and controlled input — so agents can understand and act on the computer, but only within limits.

  • Windows-oriented desktop interfaceBuilt
  • Controlled input operationsBuilt
  • Desktop awarenessDev
  • Screen-context understandingDev
  • Screenshot-based analysisDev
  • Focus and idle-state awarenessDev
  • Media awarenessDev
  • Computer interaction & task executionDev
  • Human approval before sensitive actionsDev
02

Persistent memory

Structured long-term memory on a SQLite foundation, so context and task history carry across workflows instead of resetting every conversation.

  • SQLite-based memory foundationBuilt
  • Structured long-term memoryDev
  • Relevant context retrievalDev
  • Task historyDev
  • Knowledge organisationDev
  • Continuity across workflowsDev
03

Communication

Talking to Alef AI from the phone: a Telegram bot interface today, with voice-first interaction as the direction of travel.

  • Telegram bot interfaceDev
  • Telegram text interactionDev
  • Telegram voice-note processingDev
  • Notifications and task reportsDev
  • Human approval requestsDev
  • Voice-first interactionFuture
04

Skills & playbooks

Reusable, text-based playbooks that describe how work gets done. 19 exist in the development environment — they are working material, not 19 validated commercial capabilities.

  • 19 text-based playbooks (development environment)Built
  • Structured workflows & task decompositionDev
  • Research and operational proceduresDev
  • Agent collaborationDev
  • Quality checks & execution policiesDev
04Security & governance

Capable, but controlled.

The engineering effort goes as much into limits as into abilities. Every agent works inside permissions, policies and budgets — and a human can stop everything at once.

Emergency stop
One control halts every agent.
  1. 01
    Permission-based execution

    Agents act only within the permissions they are given.

  2. 02
    Owner-only temporary access

    Access is granted by the owner, for a limited time.

  3. 03
    Prompt-injection defences

    Untrusted content is treated as data, not instructions.

  4. 04
    Per-agent policies

    Each agent has its own rules about what it may do.

  5. 05
    Agent-level budgets

    Spending and usage are capped per agent.

  6. 06
    Approval requirements

    Consequential actions wait for a human yes.

  7. 07
    Sensitive-action restrictions

    Some actions are off-limits by default.

  8. 08
    Audit logs

    What each agent did, and when, is recorded.

  9. 09
    Backup mechanisms

    State can be restored if something goes wrong.

05Research & development

Where the work is going next.

Development objectives — not completed features or promised dates.

Full roadmap
  1. 01Strengthening real multi-agent parallel execution
  2. 02Improving peer review and result verification
  3. 03Improving recovery and failure handling
  4. 04Validating Agent Town collaboration
  5. 05Improving approval and permission workflows
Product vision

A one-person business, run with a team of agents — and the owner always in charge.

Interested in the research, a collaboration, or following the build? Get in touch.