Boost Business Automation with AI Agents and MCP: Real-World Insights

Introduction

This article shares the practical journey of applying AI Agent technology to two real business scenarios – an "Intelligent Broadcast Assistant" and a "Batch Task Creation" workflow. It argues that deep integration of AI agents with traditional engineering systems, rather than trying to replace them, is the most effective way to improve business efficiency and deliver value.

Model Context Protocol (MCP)

MCP (Model Context Protocol) defines a set of standard rules that allow large language models to safely and orderly access external resources, greatly extending the capability boundary of agents. MCP consists of two roles: the MCP Client, which calls services, and the MCP Server, which provides capabilities and tools. Communication can be via STDIO (local, safe), SSE (HTTP server‑sent events), or StreamableHttp (supports reconnection and data loss detection).

Playwright‑MCP Tool Set

Playwright‑MCP is an MCP server that provides browser automation capabilities. It offers a rich set of tools such as browser_click, browser_navigate, browser_snapshot, browser_type, and many others for interacting with web pages, handling tabs, taking screenshots, and even generating PDFs.

Setup Steps

Install a compatible AI chat client (e.g., IdeaLab or Cherry Studio).

Add an MCP Server in the client settings, choosing STDIO mode and specifying the command npx @playwright/[email protected] -y --headless (or the latest version).

Configure the model service (e.g., OpenAI or Anthropic) with the appropriate API key and model ID that supports tool calling.

Create an agent with a system prompt that outlines the business workflow, required tools, and output format.

Define schedule, model parameters, and trigger keywords for each scenario.

Use Case 1: Intelligent Broadcast Assistant

The assistant monitors daily reports, detects anomalies (e.g., metric < 10% or < 30%), and sends targeted DingTalk notifications. The prompt includes steps for opening the report page, locating the metric, evaluating thresholds, and formatting the alert. The agent uses Playwright‑MCP to navigate, capture snapshots, and extract data without human intervention.

Use Case 2: Batch Task Creation

During promotional periods, operators need to create, modify, or pause hundreds of tasks based on an Excel file. The workflow was split:

Engineering part : Parse the Excel file, store task metadata in a relational database, and provide a keyword‑based lookup service.

Agent part : Use the keyword service to match uploaded tasks with existing ones, decide whether to create, update, or pause, and invoke custom tools (via MCP) to perform the actions on the web UI.

The agent’s responsibility was limited to the semantic matching (NL2Task) while the engineering side handled the heavy data processing, resulting in faster, more accurate outcomes.

Challenges and Solutions

Token limits: Large input (full task list) caused high token consumption; solution – split tasks and use asynchronous polling with Redis caching.

Browser window size affects snapshot quality; set a standard size (3840×2160) or let the agent request adjustment.

Prompt design must constrain the agent’s scope and enforce retry limits to avoid endless failures.

DingTalk markdown does not support large tables; keep alerts concise.

Comparison with FBI Broadcast

FBI provides simple scheduled broadcasting but cannot perform arbitrary web interactions. AI Agent + MCP offers flexible tool calling, enabling complex workflows such as navigating pages, extracting data, and performing actions, while still allowing FBI for straightforward alerts.

Conclusion

The key takeaway is to combine the strengths of AI agents and engineering: let agents handle tasks that require flexibility and tool usage, and let engineered components manage deterministic, data‑intensive operations. This hybrid approach yields efficient, reliable solutions and avoids over‑reliance on probabilistic AI for tasks better suited to traditional code.

References

StreamableHTTP protocol article: https://blog.csdn.net/zhangzhentiyes/article/details/147855601

Playwright‑MCP repository: https://github.com/microsoft/playwright-mcp

ModelScope MCP community: https://modelscope.cn/mcp

Cherry Studio: https://www.cherry-ai.com/