How to build AI agents that work: the harness around the model, tool calling, MCP, agents working together and a complete example in PHP.
An AI agent is a language model that calls tools itself to finish a task. Whether that works reliably depends at least as much on the software around the model as on the model itself. That is what this topic is about.
The basics are in What is an agent harness? and Agent harness: why the software around the model decides whether your AI agent works. How an agent uses tools is covered in What is MCP?, about the open standard for connecting AI to tools and data, and in Programmatic tool calling, where the model writes code that calls the tools. OpenAI WebSockets is about speed when an agent calls many tools.
If you want to build one yourself, Building an AI agent in PHP is a complete example with Composer, the Claude SDK, tests and an MCP server. Multi-agent AI shows why it so often goes wrong when agents hand work to each other, and Hermes Agent how an agent records its own way of working in skills.
LIKE '%Laravel%' finds the bootcamp graduate and misses the API architect. How to build a database that searches by meaning with MySQL, PHP and OpenAI: tested code, measured scores, RAG with GPT, and why the top score is not the best candidate.
How do you build an AI agent in PHP, without Laravel? A customer service agent for a web shop with Composer, the official Claude SDK and a home-grown framework of 280 lines: Claude chooses, PHP decides, a human approves refunds. With tests and an MCP server for Claude Code.
Several AI agents passing work to each other: when it pays off, why research says it fails so often, what A2A and MCP do and do not cover, and what a good handoff looks like in PHP.
MCP is the open standard that lets AI applications discover and use tools and data. How it works, what the July 2026 specification changes and why securing integrations is still your job.
Agent = model + harness. What an agent harness is, why the same model can score much higher in a different harness, and how guides and sensors steer an agent.
Hermes Agent doesn't learn inside the model: after hard tasks it writes its own skill files. How skills, memory and the risks work.
OpenAI's Responses API has a WebSocket mode. What it does, why agents with many tool calls get up to 40% faster, and when you can skip it.
With programmatic tool calling, the model writes code that calls tools, and intermediate results stay out of the context. Sometimes 99% fewer tokens, sometimes more expensive.
Whether an AI agent works reliably depends at least as much on the harness as on the model: context, memory, tools and error recovery. And less is often more.