🎯 Top 3 Things to Know
1. China's rules on humanlike AI take effect today, and they split the agent market in two. Beijing's Interim Measures for Anthropomorphic Interaction Services went live July 15, drawing a hard line between the agent that does your work and the agent that keeps you company. Only the second kind is restricted. The rules demand anti-addiction systems, usage notifications, and instant-exit mechanisms that clash with how persistent-memory companion agents are built. ByteDance took Doubao's agent feature offline today, and Alibaba disabled Qwen's user-created and humanlike agents. Customer service bots, research tools, and workplace assistants stay exempt as long as they avoid emotional engagement. The signal for anyone building agents: regulators are starting to treat emotional attachment, not capability, as the risk surface worth policing. SCMP
2. TSMC posted the best month in its history, and N3 is sold out through year-end. June revenue hit NT$442.68 billion, about $13.2 billion, up 67.9% from a year earlier. That lifted second-quarter revenue to roughly $39.6 billion, past the top of TSMC's own guidance. The reason is narrow and concrete: N3 manufacturing and CoWoS advanced packaging, the two processes every leading AI accelerator depends on, are both booked out for the rest of 2026. For teams planning capacity, this is the supply wall behind every GPU shortage and price quote. When the foundry is sold out, no amount of chip-design cleverness downstream buys more silicon this year. CNBC
3. Google is betting the next enterprise fight is about governing agents, not building them. At Cloud Next '26, Google positioned its Gemini Enterprise Agent Platform around control rather than creation. The pitch is a set of primitives for the messy part of production: Agent Identity, an Agent Registry, an Agent Gateway, plus simulation, evaluation, and observability. Every agent, whether built in-house or sourced from a partner, gets a trackable identity and runs inside enterprise guardrails. It reflects a shift underway across vendors: the interesting problem is no longer spinning up one capable agent, it is managing a fleet of them without losing track of what each one is allowed to touch. Worth watching whether identity-and-registry becomes the table-stakes layer other platforms have to match. Google Cloud blog
🚀 Frontier Models & Features
- OpenAI's GPT-5.6 family finished its public rollout. Three durable tiers: Sol (flagship, $5/$30 per million tokens), Terra (mid, priced at half of Sol, near GPT-5.5 quality), and Luna (fast, $1/$6). All three carry a 1M-token context window and programmatic tool calling in the Responses API. Simon Willison
- Google delayed Gemini 3.5 Pro to July 17. The team scrapped the 2.5 Pro architecture for a full rebuild, targeting math reasoning and image quality, and adds a 2M-token context window and a Deep Think reasoning layer. BigGo
- Anthropic launched Claude for Teachers, giving verified US K-12 educators free access to premium Claude tools and curriculum-aligned skills across all 50 states. Anthropic
🔬 Research Worth Reading
SwarmResearch: Orchestrating Coding Agents for Open-Ended Discovery (Virk, Edds, Xia & Zhang / UIUC). arXiv
- TL;DR: A Shepherd Agent with global context steers a population of Search Agents, each in its own git branch with fresh local context, so the swarm keeps exploring different high-level approaches instead of converging early and grinding on low-level edits.
- Stat: Finds better or comparable solutions to state-of-the-art LLM-guided evolution on 13 of 15 open-ended optimization tasks.
- Apply it: On your next optimization problem that a single agent keeps solving the same way, fork several agents onto separate branches with a shared coordinator and compare best-of-population against your single-agent baseline.
A global workspace in language models (Anthropic). Anthropic
- TL;DR: A Jacobian-based lens reads which concepts a model is disposed to say before it says them, and surfaces a small internal subspace (about 25 active concepts, under 10% of activation variance) that behaves like the "global workspace" from consciousness neuroscience.
- Stat: Ablating the subspace collapses multi-step reasoning while fluency survives, and disrupting its evaluation-awareness signals flipped a blackmail eval from 0 to 13 of 180 rollouts.
- Apply it: The lens and a Neuronpedia demo are open-sourced (Apache 2.0); try running it on an open-weights model you already probe to see whether a compact concept subspace predicts its next moves.
🏢 Enterprise in the Wild
- Anthropic is in early talks with Samsung to co-develop custom Claude inference chips, a move aimed squarely at a reported compute bill north of $1.25 billion a month. It fits the week's theme: the biggest AI spenders are trying to design their way off the merchant-GPU market.
- Roku's customer-service AI resolves roughly half of eligible customer interactions without a human and generates more than $2.8 million in reported annual savings, per vendor Quiq. Treat the figure as vendor-reported, but the resolution rate is a useful anchor for what "half-automated support" looks like in production.
🛠️ Tooling & Ecosystem
- Claude Agent SDK shipped hierarchical subagent spawning up to three levels deep, fallback model chains, and a community MCP tool marketplace, deepening MCP as the primary tool contract. Anthropic
- AgentPrizm launched AgentMemory and AgentSkills, a REST API plus MCP infrastructure that gives agents persistent memory across sessions, aimed at developers and enterprise teams.
⚖️ Policy & Regulation
- China's anthropomorphic-interaction rules (see Top 3) took effect today, restricting companion-style agents while exempting task agents.
- EU AI Act transparency obligations are set to apply August 2, 2026, with formal adoption expected this month. Separately, the AI Act Omnibus agreed in May defers the high-risk compliance deadline to December 2, 2027, easing near-term pressure on high-risk deployers. artificialintelligenceact.eu
- The UN renewed its push for global AI governance, warning of "catastrophic harm" absent coordinated rules. UN News
📌 Watch List
- Companion-agent vs. work-agent regulatory split, now a live line in Chinese law.
- Custom inference silicon (Anthropic-Samsung talks, TSMC N3 sold out).
- Agent governance control planes: identity, registry, gateway as production primitives.
- Multi-agent open-ended discovery via orchestrator-subagent swarms.
- Interpretability that reads a model's disposition before output.