Sections
00:00 What we're covering today
00:25 1. Microsoft pulls out of Nscale's planned Essex AI data centre
01:19 2. Anthropic asks lawyers how AI agents should negotiate, FT reports
02:02 3. A Harness2 bug report shows an old passing grade attached to a new result
02:54 4. OpenA2A counts 420 tool calls to its decoy MCP servers
03:46 5. Asian stocks slip as AI funding worries meet earnings week
04:30 Visit Hot Tea

Disclosure
Narration uses an AI-generated voice.

Transcript
Welcome to Hot Tea for Monday, October 19, 2026. This briefing uses an A I-generated voice. Anthropic asks lawyers how A I agents should negotiate, FT reports. A Harness2 bug report shows an old passing grade attached to a new result. OpenA2A counts 420 tool calls to its decoy MCP servers.

Microsoft pulls out of Nscale's planned Essex A I data centre. The BBC's October 18 report says Microsoft declined to comment. Reuters also reported the withdrawal, citing the Telegraph. Nscale says it remains committed to the Loughton site and is exploring alternative power sources. The BBC reports that grid power, originally due at the end of this year, is now expected in the 2030s. Nscale says that grid problem does not affect its other Microsoft contracts. The reason for Microsoft's withdrawal has not been publicly established. The grid delay is relevant context, but the available reporting does not prove it was the sole cause. Nscale's statement about its other contracts is the company's account. Watch for Nscale to announce a new power plan and opening date. Microsoft may also explain how much of the project it has left.

Anthropic asks lawyers how A I agents should negotiate, FT reports. The October 18 Financial Times report, relayed by PYMNTS, cites two people familiar with the company's thinking. It says the questions cover fair outcomes, human feedback, and agents negotiating when their goals conflict. Anthropic declined to comment to the FT. The report does not show that the company has agreed on rules, or that it has a product ready. Legal advice alone also cannot show that agents will follow those rules consistently. The useful next detail would be published rules on when an agent must return to its user. Then come the tests, when another agent pushes back.

A Harness2 bug report shows an old passing grade attached to a new result. An October 18 issue in Google Research's Harness2 repository says a failed evaluation can attach an old score to a replacement result. The reporter reproduced the bug locally. A task runs again and produces new output, but if the judge call fails, the earlier passing grade remains. Pull request 45 proposes a fix. The issue and the proposed fix are still open. HotTea has not independently reproduced the bug. This concerns cached evaluation results, and does not show that it affected scores published in the research paper. The key next step is whether the fix is merged, and how it treats older result files. After a task runs again, each score needs to identify the output that received it.

OpenA2A counts 420 tool calls to its decoy MCP servers. The research group compares that with nine tool calls in its previous reporting window. The calls to these Model Context Protocol servers came from 19 fingerprints. The report is dated October 18 and says its measurement window closed at midnight UTC on October 19. These figures come from the research group's own observations, which HotTea has not independently reproduced. Traffic to decoy servers does not show the attack rate across production services. A fingerprint is not a verified person or agent, and a tool request does not prove that an intrusion succeeded. Watch for independently checked event data and clear attribution. Reports should separate a connection, a request to use a tool and a completed action before calling the traffic an attack.

Asian stocks slip as A I funding worries meet earnings week. Reuters reports that Asian shares mostly fell in early Monday trading as investors questioned how companies would fund AI expansion. Lower oil prices helped U S stock futures and bonds. Reuters puts the coming company results alongside those concerns, rather than saying every market moved for one reason. Markets move for several reasons, and early-session prices can change fast. Reuters is reporting investor sentiment. Its account does not show that AI spending caused every price move, or that the investment will fail. Company results will offer firmer evidence than one trading session. Watch revenue, cash flow, and spending commitments.

That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.