NEW Evals & testing
Your eval score might be lying: the grader can be the bug
A score jumped from 66.1% to 87.9%, and part of that came from fixing broken graders. How to check the checker, with a checklist to copy.
// AIS
AI, broken down. Diagrams included.
Detailed breakdowns of what's new in AI and data architecture: agents, guardrails, pipelines and the tools people actually use. Every post has a real end-to-end diagram you can click through, and links to the original sources so you can check it yourself.
End-to-end diagramShort version firstSources linked
All breakdowns
Agents & reliability
The agent said done. The database said no.
Microsoft's ThinkingBox grades agents on the records they leave behind. 93.89% solved at least once, 13.41% every time.
Scale + implement
Not every change needs a human. Some need two.
How the Claude Code team sorts every change by risk: 0, 1 or 2 human approvers, verified before anyone looks.
AI coding agents
Agents are just markdown files now
What changed in Claude Code, Codex, Cursor and Antigravity this fall, and the one line in an agent file that deserves real thought.
Agents & guardrails
You approved the tool. Then the tool changed.
An approval is a moment. A tool is a stream. The Deadbugz MCP case and the defenses that keep up with it.
Tools in practice
Scheduled agents fail quietly. Silence isn't success.
When an agent runs while you sleep, a failed run can look exactly like a quiet day. Anthropic's playbook for unattended runs.
Architecture
Prompt-only vs pipeline + AI: code computes, AI explains
Upload a spreadsheet to a chat and the totals can change between runs. Here's the pipeline that keeps the numbers the same, with the AI still doing the part it's great at.
Architecture
BI-ready vs AI-ready: does your architecture have a way back?
A dashboard and an AI model can read the same data. They need different architecture. The difference is one arrow, and it changes three things.
Nothing in this topic yet. The next one is on the way.