Ephemeris.
Six dispatches
Today the work keeps sliding toward the agents — one API that picks its own models, a coding agent you drive from a phone, a research assistant that runs for a day — and the lines teams are drawing to keep them in bounds.
One endpoint that conducts the whole bench.
Sakana's Fugu shows one OpenAI-compatible endpoint and, behind it, hands each step to whichever model or agent fits — selecting, delegating, verifying and synthesizing on its own. Fugu Ultra lands shoulder-to-shoulder with Fable 5 and Mythos Preview on engineering and science benchmarks, without binding you to a single provider. Roughly 500 beta developers pointed it at code review, research, cybersecurity and patent work.
POST https://api.sakana.ai/v1/chat/completions
└─ route → reasoning model (plan)
└─ route → code model + agent (execute)
└─ verify → second model (check)
└─ synthesize → one answer
Your cloud agents now run from a phone.
Cursor 3.9's iOS app lets you kick off a cloud coding agent by voice and then walk away. It's a public beta on every paid plan — and it also reaches back to the agents already running on your desktop.
- 1Dispatch by voice — describe the task, launch a cloud agent, close the app.
- 2Remote-control the desktop — steer the agents running back on your machine.
- 3Watch it work — Live Activities and push notifications track status live.
- 4Review and merge — read diffs, screenshots and logs, then merge the PR from the phone.
A shared scale for how bad a jailbreak is.
Anthropic and its Glasswing partners — Amazon, Microsoft, Google and others — published a Cyber Jailbreak Severity scale, 0 to 4, that scores an attack on capability gain, breadth, ease of weaponization and discoverability, with severity climbing exponentially rather than by steps. The same post lays out how Fable 5 sorts cyber requests into tiers, blocking exploit development and malware while letting patch management and incident response through. It hands security teams a common vocabulary for arguing about risk.
Sort the bots into three lanes.
Cloudflare now lets any site — free plans included — treat AI crawlers as three separate things, and allow, charge or block each on its own instead of turning every bot away at once.
Search
Indexes your pages so people can still find them.
default: allowAgent
Real-time automation acting for a person — think a browsing assistant.
your callTraining
Scrapers collecting data to train or fine-tune a model.
block or chargeFrom 15 Sep 2026, new domains default to blocking Training and Agent bots on ad-supported pages while letting Search through. Opt out before the date if that's not your policy.
Hand it the question, get a day of research back.
Marlin is Sakana's first commercial product: an autonomous research assistant that runs unattended, forming hypotheses, gathering evidence and checking its own findings before it writes the report. It's aimed at strategy desks in finance, consulting and think tanks. After a 300-person closed beta, it's self-serve now, with a free monthly tier and paid plans above it.
A linter for the rules ESLint can't see.
Structural conventions, enforced Vercel Labs' konsistent is a CLI linter for the structural rules ordinary linters miss — that a file exports the function it should, that a class implements the required type, that a folder holds the files it must. You declare the conventions in a project file and run it in CI, giving human developers and coding agents the same explicit map of how the project is built. It already backs the AI SDK and Chat SDK.
{ "rules": {
"exports": "every route exports GET",
"implements": "Service → BaseService",
"contains": "feature/ has index.ts"
} }
$ pnpm konsistent → check in CI
That's today.
Six dispatches for Saturday, 4 July 2026 — a day that leaned hard on agents, and on the guardrails going up around them.