Category
Agents, tools and multi-step tasks.
AGENTS.md, Agent Skills and MCP are not rival layers. They differ by when they load, what they cost in context and whether anything enforces them. A source-checked decision table and a worked example.
Prompt, context and harness engineering are three scopes of one job. A dated, source-checked timeline of who said what, where the sources disagree, and a way to decide which fix an agent mistake needs.
Cursor, Claude Code and Codex read different instruction files. How AGENTS.md gives them one shared rulebook, and the Claude Code gotcha to avoid.
Several AI systems report perfect IMO 2026 scores. Which were marked by the organisers, which by AI agents, and how one setting moved a score by 14 marks.
The 2026-07-28 MCP spec removes sessions and the handshake. What changed, what breaks, what is deprecated, and how to migrate, checked against the changelog.
A harness is everything in an AI agent except the model. Here is what a harness engineer actually builds (tools, context, memory, permissions, budgets, evals) and how to do it well.
Underneath the hype, an AI agent is a loop: a model chooses a tool, the tool runs, the result goes back in. Here is that loop in detail, the engineering that makes it safe, and how to decide whether you need an agent at all.
A short conversation is usually enough to work out whether the idea needs a prototype, a product or a different approach entirely.