AI Agents Go to Work | AI Field Notes #62

AnalysisAnyone who reshaped a roadmap around Gemini 3.5 Pro is waiting on a model that keeps slipping. Google showed it at its I/O conference on May 19 and promised general availability in June; entering the second week of July it sits in limited preview on Vertex AI (Google's platform for building with its models), with no firm date, no published benchmarks, and provisional pricing floating near $15 and $60 per million input and output tokens. The 2-million-token context window, the largest in any production model, is the draw. The gap between a polished demo and a shippable frontier model is widening, and teams on Google's timeline keep paying for the wait.

AnalysisFifty AI agents read 466 million lines of government code in 20 hours, a job that would have kept a human security team busy for months. Anthropic (the AI lab behind Claude) published the Alberta case study on July 6: the province ran Claude Code across ministry systems holding tax records, procurement data, and social-services case files. The agents flagged known-bad patterns with a rules engine, then reviewed each flag and cited the exact file and line so an engineer could verify it before any patch shipped. The work that used to define a security job, reading code for weeks to clear one system, now fits inside a single overnight run.

AnalysisByteDance and Alibaba are pulling their personalized AI companion features rather than rebuild them, days before a new Chinese rule takes effect on July 15. The regulation, the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, was issued on April 10 by the Cyberspace Administration of China with four other agencies. It requires anti-addiction systems, usage notifications, and instant-exit mechanisms, demands that sit badly with agents built to hold a persistent emotional bond over time. ByteDance's Doubao, with 345 million users, gives people until October 15 to export their data; Alibaba's Qwen has offered no migration path at all. Beijing has decided a bot built to feel like a friend is a design it will not allow.

AnalysisMicrosoft cut 4,800 roles on July 6, about 2.1% of its global workforce, and told staff plainly that AI is changing how work gets done. The number lands inside a brutal year: US tech has shed roughly 154,000 jobs in 2026, and 56% of the layoff events that cite a reason name AI or automation. Oracle leads the tally at more than 21,000 cuts over twelve months, filing that AI deployment drove the reductions. Underneath runs the quieter pattern, companies posting record revenue while trimming headcount. The clearest signal is where the cuts land: junior hiring is slowing most, so the bottom rungs of the career ladder are the ones disappearing.

AnalysisSK Hynix is chasing roughly $28 billion in a US listing that would be the largest first-time share sale by a foreign company, a bet that the AI memory shortage has years left to run. The South Korean chipmaker supplies the high-bandwidth memory (HBM, the stacked memory that feeds AI accelerators) that Nvidia, Google, and Microsoft depend on. It plans to price around July 9 and trade on the Nasdaq the next day, after first-quarter revenue rose nearly 200% year over year and the stock climbed about 260% in 2026. The cash goes to new fabs in South Korea and to EUV machines, the extreme-ultraviolet tools that print the finest chip features. When the memory supplier raises this much, memory itself is the bottleneck.

AnalysisTesla put driverless robotaxis on Miami streets on July 3 with no human in the front seat from day one, the first time it has skipped the safety-monitor phase entirely. Ashok Elluswamy, Tesla's head of AI software, confirmed the fully unsupervised operation within hours. The service covers a 10-to-14-square-mile slice of western Miami-Dade and makes Miami the first driverless market outside Texas. What Tesla has not shared is any of the data that would let an outsider judge the call: no fleet size, no per-mile safety record for any unsupervised city, and no independent certification behind removing the monitor. The company is asking Miami riders to trust a camera-only system in Florida rain on its word alone.

AnalysisThe Federal Trade Commission opened a fight over who gets to tell AI companies how their models may answer. On July 1 it sought public comment on a policy statement arguing that firms which secretly bend model outputs for ideological ends could be deceiving consumers under Section 5 of the FTC Act. The same document takes aim at Colorado's Artificial Intelligence Act, saying state rules that push companies to alter truthful AI outputs are impliedly preempted, an approach President Trump directed by executive order. Comments run through July 31 under docket FTC-2026-0859. The move reframes model behavior as a federal consumer-protection matter and tries to pull a growing thicket of state AI laws back to a single national line.

AnalysisFor the first time, 169 governments sat down together to argue about who controls frontier AI, as the inaugural UN Global Dialogue on AI Governance opened in Geneva on July 6. Co-chaired by journalist Maria Ressa and AI researcher Yoshua Bengio, the two-day meeting is less about passing rules than admitting the gap: Bengio's scientific panel warned that AI is approaching or passing human ability in many areas while outrunning both the science that explains it and the governments meant to steer it. No binding treaty came out of the room. What did come out is an official acknowledgment, at the scale of the whole UN, that the technology is moving faster than anyone's ability to govern it.