The most capable agent-orchestration system Anthropic has ever shipped is one you cannot use if you write code on your own machine.
Anthropic rebuilt Claude Code's projects feature so that one conversation can scope work, delegate it, and run it across parallel threads — then limited the rollout to select Pro and Max subscribers using cloud sessions who have no existing projects. Everyone else waits. That includes the ZDNET writer who tested it, because his Claude Code work happens in a terminal on local files.
Last week the agent story was consumer-facing: Siri gaining agentic actions, the Astra rollout widening. This week it moved onto payroll.
The agent story moved from consumer conveniences to payroll work, while the best orchestration remains cloud-only for most developers.
Cloud users get the version Anthropic is advertising. As the company describes it, "Claude scopes the request, delegates the work, coordinates parallel threads, reviews the outputs, and assembles the finished result." Each thread is a named cloud session with its own branch and copy of the repo, and the coordinator routes new work into threads that already exist. Anthropic's own framing for how you should operate it: brief Claude "the way you'd brief a chief of staff."
The pitch is control during flight, not just after it. Because threads are named, you can drop into a running one and redirect it instead of waiting for a bad result to finish burning tokens. Threads draw from and add to shared memory, so prompt engineering should shrink over time. And it keeps running after you close the laptop, with mobile check-ins coming.
Now the parts Anthropic did not lead with. Memory is barely documented — no apparent way to inspect what was kept, correct it, export it, or learn how long it is retained. "Projects can run several threads at once, and each one is a full Claude Code session. Because of this, projects reach usage limits faster." Conflicts between overlapping threads are not prevented; they surface later as merge conflicts you resolve by hand. And this is not really autonomous: threads appear to start when you press Start on each one Claude recommends.
Local is coming, per Anthropic, "very soon." Nothing in the announcement removes terminal or CLI functionality. But "soon" is doing a lot of work in a week when the same vendor published a number that argues for urgency.
Anthropic says Claude now leads 26% of its own AI research and development work, up from 1% in March, and collaborates on more than 90%. Read that twice. The company selling agentic orchestration is running its own R&D through it, and the share where the model takes the lead went up 25 points in roughly six months.
That is why the second Anthropic item matters more than it looks. The company outlined metrics for tracking frontier AI development: how much R&D is done by AI, how well agents are overseen, and how compute is allocated. If 26% of your research is agent-led, oversight stops being a governance slide and becomes an operating line item.
Meanwhile the hardware side stopped being a hobby. Asus launched the Ascent QN10, a 1.65-pound mini PC it calls the world's first agentic AI mini PC at 80 TOPS, using Qualcomm's X2 Elite — up from 45 TOPS on the prior generation. It runs Windows 11 Copilot+ and is explicitly positioned to run agents like OpenClaw or Hermes Agent locally. The 16GB model is $1,349; the 32GB is $1,699 and is already selling out across Amazon, Newegg, and Asus's own store.
Note the architecture wrinkle: Qualcomm silicon means Arm, not the x86 you associate with Intel and AMD. For a lot of local agent tooling, that is still a compatibility question, not a spec-sheet footnote.
Perplexity went further with Nvidia. Portable Computer runs its entire agent stack — model, harness, inference engine, tools, connectors, sandbox — on hardware you own, with work consuming no billing credits and every task starting on-device by default. It asks permission before sending any individual step to a cloud model. Launch support is Nvidia DGX Spark and Linux machines with RTX GPUs; Windows follows in September. The floor is 24GB of VRAM, roughly an RTX 3090 or newer.
The interesting detail is the harness research, not the demos. Perplexity found that models advertising 260,000-token context windows start struggling past 100,000, so it built a deliberately minimal harness with on-demand skills and converted token-hungry connectors into compact command-line tools. If the OS-level sandbox is unavailable, the harness disables itself rather than run unprotected tools.
Dell has been selling the enterprise version of this since May. Its Deskside Agentic AI line handles models from 30 billion to 1 trillion parameters at the desk, with Nvidia's NemoClaw stack and OpenShell sandbox spanning workstations up to PowerEdge servers. Dell claims break-even against public cloud APIs in as little as three months and up to 87% lower spend over two years — and notes that over half of agentic workflows run on open-weight models.
So the contradiction resolves into something simpler than either camp admits. Local inference is no longer an ideology; it is a placement decision with real economics for high-volume, always-on work and no advantage for occasional work. That framing comes from a longer analysis of desk-side agent adoption, and it matches the hardware: the memory threshold moved, so trillion-parameter inference at a desk became a product rather than a demo.
Which makes Anthropic's cloud-only launch window more revealing than a normal staged rollout. The company with the best orchestration story can't yet run it where a growing share of professional work is heading. Threads run in the cloud today; running on your machine alongside your local tools and code and behind your network is coming very soon.
Should you care? Yes, but not for the reason the launch posts suggest. The orchestration layer is real and it changes how a solo developer or small team sequences work — if you already live in cloud sessions. If you don't, you got a roadmap and a reminder that the terminal remains the fallback.
Should you buy? If you have a recurring, high-volume agent workload and data you would rather not upload, the $1,349 Asus or a 24GB RTX machine running Portable Computer is now a defensible purchase rather than an experiment. If your agent use is occasional, the cloud still wins on every axis except privacy.
The oversight metrics are the item to watch. When a lab reports that its model leads a quarter of its own research, the question stops being whether agents work and becomes who is watching the 26%. Anthropic proposing the measurement framework is a good sign and a self-interested one at the same time — the same company is asking you to trust its numbers about its own agents.