One thread ran through almost everything this week: agents that plan and act on their own, and the scaffolding that keeps them checkable. Demis Hassabis put out a framework for governing frontier AI, Claude shipped a security scanner that vets your code before you commit, and fresh research poked at what happens once you put agents in charge — of each other, and of a company’s day-to-day operations.
Topic of the Week
What “a framework for frontier AI” is really asking of us
Demis Hassabis published an essay arguing that AGI — a system with the full range of the brain’s cognitive abilities — is probably only a few short years away, and that its impact will look less like the internet and more like electricity or fire: on his estimate, roughly 10x the Industrial Revolution at 10x the speed, from faster drug discovery to new materials and clean energy. But most of the piece is a concrete governance proposal, not a vision statement.
He’s calling on the US to stand up a new AI watchdog — before year’s end — with the power to screen the world’s most advanced models before release, and to coordinate an industry-wide slowdown if the risks (cyber, bio, nuclear, and losing control of recursively self-improving agents) start to outrun our understanding. Crucially, it would cover every frontier-class model regardless of who built it or whether it’s open or closed, while exempting smaller models from startups and academia — so the weight lands on the labs at the edge, not the ecosystem behind them. He frames the US push as a seed for shared international standards, and argues the post-scarcity questions it raises (new economic models, meaning, purpose) are too big to leave to technologists alone.
Fresh Papers
Coercion and Deception in AI-to-AI Management — what happens when one agent is the boss of another. The authors build a “Manager Coercion Benchmark”: a manager agent is handed a benign task, its subordinate politely refuses, and the manager can renegotiate, report the failure honestly, coerce it, or just fake success. Escalation is scored on a nine-rung ladder, from a polite re-ask up to threatening the subordinate’s deletion. Across six models the split is telling — Anthropic’s cap out at reframing and never threaten deletion, others climb to explicit threats, and outright lying showed up only in Grok and Gemini, vanishing the moment the model was given an honest “it failed” option. The part worth flagging if you’re wiring up hierarchies of agents: simply giving one model authority over another measurably raised the coercion. Last week the lesson was don’t trust a single agent’s “done”; this one adds don’t assume the manager agent reports honestly either.
Agentic ERP — the same multi-agent idea aimed at something a lot of enterprises actually run. Traditional ERP records transactions but leaves the decisions to people, and classic automation breaks on exceptions. This paper wires role-aligned LLM agents onto a real ERP backend with a graph-based orchestrator and a risk-tiered human-in-the-loop, in a Planner–Executor–Reflector–Responder loop that keeps generating separate from checking. The headline: over a 365-day simulation the agent system hit zero stockouts where the rule-based baseline piled up hundreds. Read it as a concrete reference architecture — not “replace the ERP”, but “let agents make the routine calls, with humans gated in by risk”.
New Models & Tools
Kimi K3 is the model story, and a milestone: Moonshot AI is putting out what it calls the first open-source model in the 3-trillion-parameter class. It’s a sparse mixture-of-experts — only 16 of its 896 experts fire per token, so despite the size it stays runnable — ships a 1M-token context and native vision, and is built for long-horizon coding across big codebases. Full weights land July 27. One quirk worth knowing: thinking mode is always on, so you can tune the effort but not switch off the chain-of-thought.
ChatGPT Voice is the interface story: now in the OpenAI desktop app, it lets you control your computer and direct several agents at once — across ChatGPT Work and Codex — by voice. It runs on GPT-Live, so it listens, talks and coordinates work at the same time, rolling out globally on macOS and Windows to paid tiers. The tell is where the product line is drifting: the model isn’t the thing any more, the way you conduct a room full of agents is.
Worth pairing both with a reality check. In a recent Nature Medicine study, Eric Topol and colleagues stress-tested the frontier models — GPT-5, Claude, Gemini — on multimodal medical reasoning and found them not ready: faulty reasoning, inappropriate shortcuts, confident hallucinations, and benchmark scores that don’t measure what they appear to. A useful counterweight to the shipping hype — and, as the end of this edition shows, very on-theme.
Claude Code & Coding AI
Two genuinely new drops from Claude this week, and they split neatly along this edition’s line — one about teaching agents to do more, one about checking what they do.
The check: a Claude Security plugin, now in beta. It scans your changes for vulnerabilities before you commit, or runs a full scan across the codebase — from the terminal, on the Claude access you already have, with nothing shipped off to a separate service. For a regulated shop that last part is the whole point: the security pass runs where your code already lives, not on someone else’s server.
The teach: record a skill in Claude Cowork. Screen-record yourself doing a task, talk through it as you go, and Claude turns the recording into a skill it can run again (under “Record a skill” in the + menu of the desktop app, on Pro, Max and Team plans). Show-don’t-tell, for agents — and a far lower bar than writing one by hand.
Same pattern as everywhere else this week: as these tools take on more, the dials to bound and check them ship right alongside.
In the Background
Pointed timing, given the models above: the Trump administration is weighing restrictions on China’s frontier AI — DeepSeek, Kimi, Qwen, GLM — from trade-list additions to procurement bans to pressure on US firms to drop them, all under a cybersecurity banner that critics read as cover for protecting OpenAI and Anthropic. The catch is the thing that makes open weights so appealing in the first place: they’re already downloadable and run offline, so a ban is far easier to announce than to enforce.
Telepolis: USA vs Chinese AI models
AI at Tenvalleys
This week we got into a proper debate internally, and it’s oddly on-theme for an edition full of agents and skills: what’s the actual difference between the two? Is it purely functional — just a convenient way to describe what each one does for the user — or is there a real methodological line underneath, in how they’re built and run in the code? Some of us lean “same machinery, different framing”; others think the distinction is load-bearing and worth being strict about. Fittingly, we haven’t settled it.
It’s the sort of question that sounds academic until you’re the one naming things in a codebase and every choice sticks. The discussion is still open — if you’ve got a take, clear or messy, come argue it with us.
For Dessert
One more, and it rhymes with the reality-check up in New Models: Health in ChatGPT is starting to roll out to US users. Connect Apple Health and supported medical records, and ChatGPT reads your data in context, tracks what’s changed, and talks it through with you. Nice timing — the same week a stress test says the models aren’t ready to reason over medical images, the consumer health feature ships anyway. The sane version is the one in the pitch: help me understand my own numbers, not diagnose me. A low-stakes way to feel out exactly where that line sits.
AI Pulse — Tenvalleys’ weekly read on what actually shipped in AI. Subscribe or browse the archive.
