Agentation – Visual UI Annotation for AI Coding Agents
Click a UI element, add a note, and the agent gets the CSS selector, source file path, React component tree, and computed styles instead of a vague description like "the blue button."
AI Daily Recap
5 articles · 3 categories
The finishable daily brief
What happened in AI — Jul 11, 2026
Saturday, Jul 11, 2026
5 articles · 3 categories
read top to bottom · then stop
In 30 seconds
- Vasari and HackerRank's hiring agent both converged on the same judge recipe: rubric to prompt, validated against human labels before trusting it in production.
- Agentation turns a clicked UI element into a CSS selector, source file path, and component tree an agent can grep directly.
- AgentKindergarten tunnels a coding agent's terminal to your phone with no inbound firewall rules, plus auto-flagging when a session stalls.
- A naive Wikipedia fetch costs 68,240 tokens; a new stealth-browser MCP cut a blocked page down to about 700.
A quiet day for launches, but a consistent thread in what builders shared: tooling to make agents auditable rather than just capable. Two separate builders arrived at the same recipe for a trustworthy LLM judge — a human-audited rubric scored by a prompt, validated against human labels before it's trusted — while coding-agent tooling focused on giving agents richer context and letting builders supervise them remotely instead of blindly.
One more data point on cost: a naive Wikipedia fetch runs past 68,000 tokens, and today's fix was purpose-built extraction, not a bigger context window.
Coding-Agent Tooling 2 items
Two new tools treat the coding agent itself as fixed and build the tooling around it: one turns visual UI feedback into machine-readable context, the other streams and remotely supervises an agent's terminal session.
AgentKindergarten – daycare for your AI coding agents
A daemon-relay-client setup tunnels a coding agent's terminal and auto-detected dev-server ports to a phone or browser with no inbound firewall rules, plus a report-card grader that flags stalls and errors automatically.
Building Trustworthy LLM Judges 2 items
Two unrelated builders landed on the same recipe for a judge you can actually trust: a human-auditable rubric turned into a score by a prompt, only relied on after checking it against human-labeled ground truth.
I built a free tool to evaluate AI agent outputs (human labels and LLM judges)
Vasari has reviewers label agent traces pass/fail with error codes, then validates any LLM judge against those human labels with a confusion matrix and Cohen's kappa before it's trusted on unlabeled production traffic.
Inside HackerRank's LLM-based Hiring Agent
The pipeline scores resumes on a fixed rubric (up to 35 of its points for open-source contributions, excluding a candidate's own repos) computed almost entirely by prompt templates rather than hardcoded logic — testing found its startup-founder bonus barely moved scores.
Context & Tool-Use Economics 1 item
Naive page fetching can cost an agent tens of thousands of tokens per page; today's fix on display was extraction tooling purpose-built to strip that overhead before it ever reaches the model.
One Wikipedia page costs your AI agent 68,000 tokens
An average Wikipedia article runs 68,240 raw-HTML tokens to fetch, and Claude Code's own web tool falls back to near-empty results on JS-heavy or anti-bot pages — prompting an open-source stealth-browser MCP that cut one blocked page's cost from a 403 error to about 700 tokens.
You are caught up for this edition