Blog

Technical writing on agent frameworks, MCP servers, and autonomous agent infrastructure.

A single tall rounded rectangle outlined in orange standing on a thin horizontal line, alone on a black ground

Is Claude Code Auto Mode Safe? What the Classifier Misses

Claude Code auto mode blocked 89% of dangerous commands testers missed. Anthropic's own docs say the classifier never reads tool output. That gap is the answer.

agents security11 min read

Three agent runs drawn as horizontal tracks: one stopping at a human, one pausing at an approval gate, one continuing past an empty chair

AI Agent Autonomy Levels: Copilot, Supervised, Autonomous

Three AI agent autonomy levels, defined by one test: who approves the next step, and what happens when nobody answers. Six tools placed on the scale, with sources.

taxonomy agents8 min read

A bar chart on a black ground where two short outlined bars mark the provider outages and a taller solid orange bar marks the agent tool's longer degraded window

AI Outage September 2026: ChatGPT, Claude and Grok

ChatGPT, Claude and Grok all broke on September 3, 2026. A UTC timeline from four status pages, and why Cursor stayed degraded longer than any lab it runs on.

agents architecture9 min read

A tool call leaving a model loop, crossing an MCP client, and arriving at an MCP server over JSON-RPC

MCP vs Function Calling: Which One Do You Actually Need?

MCP vs function calling: they are not competitors. Function calling is how a model picks a tool. MCP is how a tool reaches any client. One rule tells you which you need.

mcp tool-calling8 min read

An MCP server answering a tools list request and a tool call from one agent client over two transports

What Is an MCP Server, and How Does It Work?

An MCP server exposes tools, resources and prompts to an AI app over JSON-RPC 2.0. Here is what that means in practice, and why an MCP badge in a README proves nothing.

mcp protocols9 min read