All case studies

AI Platform · Multi-Agent Systems

AgentX

A multi-agent platform that turns a natural language prompt into a production-ready full-stack application, and code a human can still maintain afterwards.

Architecture & Integrations LeadView AgentX
20+

Engineers led

Across AgentX and client product lines

50%+

Fewer integration issues

From clear service boundaries and ownership

Phase 1

Complete

Shipped and generating applications

Generated code still has to be code someone can own

AgentX generates production-ready full-stack applications from natural language prompts. It is a multi-agent system built on LangChain and FastAPI, where specialised agents handle planning, schema design, UI generation and tool orchestration rather than one model attempting everything at once.

The interesting problem is not generation. It is the gap between an abstract requirement and a concrete database schema and component tree. A prompt like “let users book appointments” implies entities, relationships, states and screens that nobody stated, and getting them wrong compounds through every file the system writes.

I built the integration layer: Figma-to-code, Supabase provisioning, and MCP tool orchestration. The constraint I designed against throughout was maintainability: generated output that a human team has to inherit is only valuable if it looks like code they would have written.

Stack

PythonFastAPILangChainMulti-agent systemsMCPSupabaseNext.jsVector databasesFigma API
The problem

What made this hard

Ambiguous requirements

Mapping abstract natural language onto concrete database schemas and UI components without inventing the wrong model.

Agent coordination

Structuring autonomous agents so their outputs compose into one coherent application rather than conflicting fragments.

Human maintainability

Generated code has to be readable and ownable by an engineering team after the agents finish.

System design

High-level architecture

Specialised agents sit behind one orchestration layer, so every decision belongs to a stage you can inspect rather than a single opaque generation.

Input layer

How a specification arrives

Natural language promptProduct intent
Figma design importFigma API

Agent orchestration

Narrow agents, one shared plan

Planning agentEntities & screens
Schema agentDatabase design
Backend codegen agentServices & APIs
UI codegen agentComponents

Tool & platform layer

Where agents act on real systems

MCP tool orchestrationModel Context Protocol
Orchestration APIPython · FastAPI
Vector storeContext retrieval

Output layer

A running app, not a code dump

Supabase provisioningAuth · DB · storage
Generated applicationFull-stack
Live previewReviewable build
Architecture

From prompt to running application

Each agent owns a narrow decision. Splitting the work this way is what makes the output reviewable, because you can point at the stage that got something wrong instead of re-rolling the whole generation.

  1. 1

    Natural language or Figma input

    The system accepts a written prompt or an existing Figma design as the starting specification.

    LangChainFigma API
  2. 2

    Planning agent

    Requirements are decomposed into entities, relationships and screens. This is the step that decides whether everything downstream is coherent.

    Multi-agentLangChain
  3. 3

    Schema design

    Concrete database schemas are derived from the plan, with relationships and constraints made explicit rather than implied.

    Schema generationSupabase
  4. 4

    Specialised codegen agents

    Separate agents generate backend services and UI components, each working within the shared plan.

    LangChainFastAPI
  5. 5

    MCP tool orchestration

    Agents call real tools through the Model Context Protocol, so they act on actual systems rather than describing intentions.

    MCPTool calling
  6. 6

    Provisioning and preview

    Supabase resources are provisioned and the generated application is assembled into a running, reviewable preview.

    SupabaseNext.js
Decisions

Why these tools, and not others

Agent framework

LangChain multi-agent

Narrow, specialised agents produce reviewable output; a single generalist agent produces something no one can debug.

Orchestration API

Python FastAPI

Keeps the platform in the ecosystem where LLM tooling lives, with a fast typed service layer around it.

Tool access

Model Context Protocol

A standard interface for agents to reach real tools, instead of bespoke glue per integration.

Backing platform

Supabase

Generated applications get real auth, storage and a database provisioned automatically rather than stubs.

Outcome

What it delivered

Phase 1 shipped

A working multi-agent platform generating full-stack applications from prompts.

Figma-to-code integration

Existing designs become a valid starting specification, not a manual re-draw.

MCP tool orchestration

Agents operate real tools through a standard protocol.

Leading 20+ engineers

Architecture standards and service boundaries that cut cross-team integration issues by over 50%.

Building something like this?

If any of the problems above look familiar, I can help, whether that's a full architecture or a second opinion on the one you have.

Open to remote and hybrid work worldwide