The frontier model becomes the supervisor

Cursor shipped an iOS app yesterday that launches always-on cloud agents from your phone, or remotely drives the agents running on your laptop, with Live Activities pinging you when one finishes or needs input. Composer 2.5 is 75% off in the app through July 5. The launch itself is incremental; the signal around it is not. Gergely Orosz, fresh out of Cursor’s offices, put it bluntly: coding agents in the cloud “are coming fast,” and local agents “will prob go away soon as I read the room.” When the person who tracks engineering practice for a living says the local-first era has a short runway, that’s worth more than the app store listing.

The same day, Cognition shipped Devin Fusion, which is the architecture answer to the same economics. Fusion is a hybrid-model harness built on two ideas: a smaller “sidekick” agent runs in parallel with a frontier agent, and the frontier agent delegates work to it while keeping ownership of planning, ambiguity, and final review. You get frontier-level judgment without paying frontier prices on every token. Cognition built it on their FrontierCode benchmark, which tries to measure whether you’d actually merge a given PR to production rather than whether the diff looks plausible, and they claim it sidesteps the usual model-routing failure modes: expensive cache misses and routers that don’t generalize past their training distribution. It’s live in Devin now. Read it next to Cursor’s cloud push and the picture is consistent: the frontier model is becoming the expensive supervisor, not the thing doing every keystroke.

Access to those frontier models, meanwhile, is turning into a political artifact. OpenAI posted a 39-minute deployment-safety preview of the full GPT-5.6 family, Sol, Terra, and Luna, expanding on the Sol preview from earlier this week, and the headline practitioners actually care about is that you can’t use it. Access runs through a government-limited program. This rhymes with Anthropic’s Mythos 5 coming back for about 100 vetted orgs while Fable 5 stays locked. Two of the three frontier labs are now gating their newest weights behind partner lists and clearance, and the question for the next year is whether “I have a model that can do X” stops meaning “I can ship X” because the model you’re cleared to call is two releases behind.

The counter-pressure comes from open weights, and it kept building. LongCat-2.0 hit the Hacker News front page, a large open mixture-of-experts model with 1.6 trillion total parameters and 48 billion active per token. The active-parameter count is the number that matters for anyone planning to actually serve it: you pay 48B of compute, not 1.6T, which puts a frontier-class open model inside reach of a serious self-hosting budget. Every gated announcement out of the closed labs makes the open-weight option look less like a fallback and more like the default for teams that want to own their stack.

On the tooling-availability front, Claude Opus 4.8 fast mode is now in preview for GitHub Copilot. It’s the same model with materially faster output token speed, aimed at the interactive loop where you’re watching the agent type and waiting on it. For day-to-day Copilot work the latency change matters more than another point on a benchmark.

If you’re building agents, Wix Engineering ran the eval most teams keep meaning to run and published it: 250 agent evals to test whether Skills beat plain docs. The answer was messier than the framing implied, which is the useful result. Structured Skills don’t uniformly win over well-written documentation; the gap depends on the task shape, and a lot of the “Skills are the new abstraction” enthusiasm runs ahead of the measured evidence. If you’re deciding whether to invest in a Skills layer or just write better docs, this is the data point to read before committing.

Last, a supply-side note that practitioners under-track: Google is limiting Meta’s use of its Gemini models, per an FT report. When the labs start fencing each other off, the downstream effect lands on anyone whose product quietly depends on a competitor’s models through a partnership that can be revoked.

What to watch: the AI Engineer World’s Fair is running in San Francisco this week, with swyx and the Latent Space crew live-blogging from the floor and a Sandbox & Platform Engineering track aimed squarely at the cloud-agent infrastructure question Cursor and Cognition just made urgent. The talks coming out of it over the next few days are where the next quarter’s tooling gets previewed.