Skip to main content
PROJECT / AI·IN DEVELOPMENT·Systems Architect & Developer

AI Workflow Platform

Autonomous Agentic Automation with MCP & Playwright

An AI-powered workflow automation system exploring autonomous agents, Model Context Protocol (MCP) servers, browser automation, and containerized cloud runtimes.

#Next.js#FastAPI#Azure#Docker#MCP#Playwright
// SYSTEM ARCHITECTURE & TOPOLOGY
01 / THE PROBLEM

The Challenge & Context

Traditional LLM agents suffer from brittle tool executions, hallucinated parameters, and lack of real-time environment feedback when executing multi-step web and cloud workflows.

02 / THE SOLUTION

Engineering Approach

Architected an agentic execution platform that utilizes Model Context Protocol (MCP) servers for standardized tool interfaces and Playwright for deterministic browser automation, running in isolated container environments.

03 / ARCHITECTURE

Components & Data Flow

A Next.js frontend interacts with a FastAPI orchestrator service. The orchestrator dispatches tasks to ephemeral Docker containers in Azure Container Apps. The containers communicate via MCP to interact with external APIs, databases, and browser sessions.

04 / ENGINEERING DECISIONS

Key Trade-offs & Decisions

DECISION_01

Model Context Protocol over Custom Function Calling

MCP provides a standardized protocol for tool discovery, client-server separation, and deterministic schema validation.

DECISION_02

Azure Container Apps with KEDA Scale-to-Zero

Workload execution is event-driven; scaling containers to zero when idle eliminates idle compute costs.

05 / IMPLEMENTATION CHALLENGES

Obstacles & How They Were Overcome

CHALLENGE: Managing Headless Browser Memory Footprints

RESOLUTION:Engineered aggressive container garbage collection and browser session recycling to prevent memory leaks during long-running tasks.

06 / OUTCOME

Measurable Impact

Reliable multi-step workflow execution with automated retry and self-healing error handling.
Modular MCP servers easily reusable across different agent architectures.
07 / WHAT I LEARNED

Field Lessons & Takeaways

Never trust raw model output for system commands; strictly validate all parameters against Zod/Pydantic schemas before tool invocation.