hellobuilder

Since Nov 5, Outlook.com, Hotmail, Live and MSN bounce part of the mail from domains with no DMARC record. Small senders too, and nothing was announced. If your app sends sign-in codes to Hotmail users, check today. The fix is one DNS record:

AI doesn't leave a blank where a number goes. It writes a plausible one. Our landing page claimed 1,000+ happy professionals. The database had 12 people. How to audit every number on your page, including the review markup only Google sees:

This week's HelloBuilder issue: my AI-written page said 50,000 headshots. The database said 440. It was also telling Google "4.8 stars from 2,500 reviews". There is no review system. What the agent found, the one file every number now has to pass, and how to check yours:

The same model can cost very different amounts depending on who serves it. Grandet compares listed API prices across official APIs, inference providers and relays, and your agent can look them up over MCP with no key. Check who handles your data first:

One developer, one agent, about 530 tests. Three of the four real bugs got past every check built for their feature. Four ways a test lies when an agent writes it, and the lab-science idea behind catching them:

AI agents now browse, compare and sign up for things. Can they use your site? Dedica runs 125 free checks the way an agent would, and gives you a fix for each failure to hand to your coding agent:

Does this output matter? Is this command safe? Is the work really done? mu hands those yes/no calls to a small judge model at 40 points in the agent loop, instead of the main model. It also catches a "done" nothing verified:

An agent hit a broken tool and retried 31,000 times overnight. $4,700 in tokens on a $400 plan. Global rate limits and spend alerts didn't catch it. The fix: a per-customer budget checked before every model call, plus a retry cap:

Your skill passes its checks. Does it trigger when an agent actually uses it? Coder Eval runs Claude Code, Codex and others in a sandbox against YAML tasks, scores what they produced, and can fail CI. Playwright for coding agents:

Did your AI rewrites work? Most people never find out. One line per change when it ships, one script that reads Search Console, four dumb verdict rules, and a control: pages you didn't touch. That's the whole grader. Ours found 35 regressions: