DEV Community

Cover image for I Built an Engineering Orchestrator for AI Coding Agents
Vishnu Choudhary
Vishnu Choudhary

Posted on

I Built an Engineering Orchestrator for AI Coding Agents

AI coding agents are getting incredibly good at writing code.

But there is another problem that becomes more obvious as projects get larger:

Who manages the engineering process around the AI?

When working with AI coding agents, we often end up repeatedly providing project context, explaining architecture, reminding the AI about conventions, reviewing generated code, and manually coordinating different parts of development.

I wanted to solve that problem.

So I built AI Engineering Orchestrator, a VS Code extension designed to add an engineering workflow layer around AI coding agents.

The idea

Instead of treating an AI coding agent as a single developer, the Orchestrator treats software development more like an engineering team.

For example:

Requirement
    ↓
Business Analyst
    ↓
Solution Architect
    ↓
UI/UX Designer
    ↓
Backend Developer
    ↓
Frontend Developer
    ↓
QA Engineer
    ↓
Security Engineer
    ↓
Code Reviewer
    ↓
Release
Enter fullscreen mode Exit fullscreen mode

Not every task needs every role.

A small bug fix might simply be:

Analyze
  ↓
Implement
  ↓
Test
  ↓
Review
Enter fullscreen mode Exit fullscreen mode

While a large feature can go through a much more complete workflow.

Roles

The extension provides specialized engineering roles such as:

  • Product Manager
  • Business Analyst
  • Solution Architect
  • Software Architect
  • Database Architect
  • UI/UX Designer
  • Frontend Developer
  • Backend Developer
  • DevOps Engineer
  • QA Engineer
  • Security Engineer
  • Performance Engineer
  • Code Reviewer
  • Technical Writer
  • Release Engineer

Custom roles can also be created.

The important part is that a role defines what the AI should focus on, what it is allowed to do, and what quality requirements it needs to follow.

Skills

Roles can use reusable skills instead of having one huge system prompt.

For example, a backend task might activate:

API Design
Validation
Authentication
Error Handling
Database Design
Testing
Security
Enter fullscreen mode Exit fullscreen mode

A frontend task could activate:

Component Design
State Management
Accessibility
Responsive Design
Performance
Testing
Enter fullscreen mode Exit fullscreen mode

Only relevant skills should be included for the current task.

The biggest focus: context optimization

One of the biggest problems with AI-assisted development is context.

A project might contain hundreds or thousands of files, but a particular task may only require five of them.

Instead of sending the entire project, the Orchestrator tries to identify the most relevant context.

For example:

Current Task
    +
Current File
    +
Relevant Function
    +
Direct Dependencies
    +
Related Tests
    +
Relevant Architecture
    +
Relevant Decisions
    +
Relevant Conventions
Enter fullscreen mode Exit fullscreen mode

The result is a smaller, focused context package.

The goal isn't simply to reduce tokens.

The goal is to provide the AI with better context per token.

Project memory

The extension maintains project knowledge so AI agents don't have to rediscover the same information repeatedly.

The project can maintain information such as:

Project Architecture
Requirements
Coding Conventions
API Contracts
Database Information
Testing Strategy
Security Rules
Tasks
Architectural Decisions
Enter fullscreen mode Exit fullscreen mode

For example:

DEC-014

Decision:
Use PostgreSQL.

Reason:
The application requires relational transactions.

Affected:
Backend
Database
Infrastructure
Enter fullscreen mode Exit fullscreen mode

A future task can retrieve the relevant decision instead of receiving the entire history of previous conversations.

Structured workflow

The Orchestrator uses workflow phases such as:

DISCOVER
   ↓
ANALYZE
   ↓
PLAN
   ↓
ARCHITECT
   ↓
DESIGN
   ↓
IMPLEMENT
   ↓
TEST
   ↓
SECURITY
   ↓
REVIEW
   ↓
DOCUMENT
   ↓
RELEASE
Enter fullscreen mode Exit fullscreen mode

The workflow is dynamic.

Irrelevant phases can be skipped.

Quality gates

AI-generated code shouldn't automatically be considered production-ready.

The Orchestrator can introduce quality gates such as:

Architecture Gate
Implementation Gate
Testing Gate
Security Gate
Code Review Gate
Documentation Gate
Enter fullscreen mode Exit fullscreen mode

For example, implementation may require:

✓ Type checking
✓ Linting
✓ Tests
✓ Acceptance criteria
Enter fullscreen mode Exit fullscreen mode

before the task is considered complete.

AI-provider agnostic

The idea is not to replace AI coding agents.

Instead, the Orchestrator is designed as a layer around them.

The architecture is provider-oriented so different coding agents can be integrated through adapters.

The long-term goal is to work with different AI coding workflows rather than locking developers into a single AI provider.

Security and guardrails

AI agents can make powerful changes, so permissions matter.

Different roles can have different capabilities.

For example:

Backend Developer

Can:
✓ Read files
✓ Create files
✓ Modify backend code
✓ Run tests

Cannot:
✕ Modify production database
✕ Expose secrets
✕ Change architecture without approval
Enter fullscreen mode Exit fullscreen mode

The extension is also designed around secret detection, controlled execution, diff review, and approval for potentially destructive operations.

What I'm trying to build

The bigger vision is:

The AI should do the coding. The Orchestrator should decide what the AI needs to know, which role it should act as, what it is allowed to do, and how the result should be validated.

This makes the extension less like another AI chatbot and more like an engineering operating system for AI coding agents.

Try it

AI Engineering Orchestrator is available for VS Code:

👉 https://marketplace.visualstudio.com/items?itemName=ai-engineering-orchestrator.ai-engineering-orchestrator

I'd love to hear from developers who are already using AI coding agents.

What is the biggest problem you face today: context, consistency, architecture, token usage, testing, or managing the AI workflow?

AI #AICoding #VSCode #DeveloperTools #SoftwareEngineering #DevTools #ArtificialIntelligence #Programming #DeveloperExperience #ClaudeCode #Codex

Top comments (0)