What I test and learn about AI coding tools such as Claude Code, AI agents and MCP, and web development with PHP.
Which AI model can you run locally on an ordinary 16 GB laptop? Qwen3.5 and Gemma 4 compared, with measured speeds, a step-by-step Ollama setup, tasks to start with and the two settings that make the most difference.
Mistral Large 4 is the first European open-weight model with a trillion parameters. What Mistral's figures do and do not say, what it costs, what hardware you need and why the licence is the biggest open question.
Jev does not write text; it answers multiple-choice questions with a probability for each answer. The principle in plain language, with PHP examples you can copy, up to Claude and Jev working together.
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.
Jev by TypeSafe AI writes no text: it picks from your list and attaches probabilities. What independent tests measure, what a decision costs, and when plain code or a trained classifier is the better choice.
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.
Cognitive debt is the gap between how your system works and how well your team understands it. What research says about it, and how to prevent it without ditching AI.
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.