Daily TEA: Why We’re Buzzing

China’s AI classrooms, open-weight politics, Europe’s telecom bill, and a practical way to test AI agents.

Jul 24, 2026

Hello, dear TEA-mates! Here is what you need to know today.

1. 🧑‍🏫 China’s AI Classroom Becomes a Global Talking Point

Entrepreneur Vaibhav Sisinty shared a video from China’s World AI Conference showing what he described as an AI blackboard, with students learning to build with AI in real time. He contrasted that approach with continuing U.S. debates over whether students should use AI at all. The post is a snapshot of a larger argument about whether schools should teach AI literacy and building skills, rather than only policing access. (Read More)

🫖 TEA For Thought: “This is so cool. China is truly leading the AI application layer. It is gradually advancing across every layer except chips and infrastructure, at least for now.”

2. 🐝 Buzz Puts People, Agents, and Code in One Workspace

Buzz has launched as an open-source workspace that places people, agents, conversations, and code behind one cryptographic identity system. Its signed, self-hosted relay records messages, patches, reviews, workflows, and approvals in one audit trail. The team says channels, threads, DMs, canvases, search, and workflows work now, while full Git hosting, mobile, push, approval gates, federation, tighter agent permissions, and a hosted option are still in progress. (Read More)

🫖 TEA For Thought: “Jack is truly one of a kind!”

3. 🧩 Startups Push Back on a Chinese Open-Weight AI Ban

Nearly 200 Silicon Valley companies, including Proton and Y Combinator, urged the Trump administration not to broadly block U.S. access to Chinese open-weight AI models. The new Little Tech Association argued for targeted safeguards instead, warning that a ban could raise costs and leave smaller builders dependent on a few large U.S. model providers. The administration had not announced a policy, while officials publicly raised concerns about alleged model distillation and export-control evasion. (Read More)

🫖 TEA For Thought: “Open source is the future. Still, it is hard not to worry about backdoors or dormant code that could create a wake-up-style access problem, like vulnerabilities found in military-device codebases. The open-source ecosystem should not be dominated by the CCP or China. Inkling, Gemma, and Lugano should try harder and receive government support to compete at the frontier.”

4. 📡 Europe’s Huawei Removal Could Cost Up to €40 Billion

Removing Huawei and other Chinese vendors from European Union telecom networks could cost operators €30 billion to €40 billion, according to a GSMA Intelligence report commissioned by seven European operator groups. That is roughly three times the European Commission’s estimate, largely because the industry study includes mobile, fixed, and transport networks. The report also projects less competition, higher equipment prices, and delayed upgrades if the proposed removal mandate is adopted. (Read More)

🫖 TEA For Thought: “It costs money to decouple, but that is better than being totally controlled by the CCP.”

5. 🧪 LangChain Wants Agents to Help Build Their Own Tests

LangChain has released an Eval Engineering Skill that helps coding agents inspect a repository and execution traces, then propose and build executable evaluations in Harbor format. The workflow asks a user to choose which abilities matter, creates the task environment and verifier together, and uses traces to reveal reward hacking or other shortcuts. LangChain argues that the useful loop is iterative: mine real failures, build an eval, improve the agent, and rerun it in a stable environment. (Read More)

🫖 TEA For Thought: “Save this and use it today.”

🛠️ Skill of the Day

The Agent Reality Check: Turn one recurring AI-agent task into a small, measurable test before you trust it with more work.

You are an AI quality lead. Help me turn one real task that an AI agent performs into a practical evaluation.

Task to evaluate: [DESCRIBE THE REAL JOB]
What success looks like: [OUTCOME, QUALITY BAR, AND DEADLINE]
Real examples or past failures: [PASTE THEM HERE]
Constraints: [WHAT MUST NOT HAPPEN, SUCH AS PRIVATE-DATA EXPOSURE, UNAPPROVED WRITES, OR EXTRA COST]

First, identify the one ability that matters most to test. Then propose three realistic test cases: a normal case, a difficult but valid case, and a failure case. For each, state the expected outcome, the evidence that would prove success, and what should trigger a human review.

Keep the test small enough to run repeatedly. Do not invent user data, results, policies, or system access. Clearly label every assumption and every missing piece of evidence.

Paste this into ChatGPT, Claude, or your tool of choice. Replace the bracketed sections with a real task and real evidence.