the proof has to survive the prompt

Zihan keeps choosing local memory over a service that charges rent on remembering. It is a small product instinct with a long shadow: the useful part of an agent is not the flourish, but the things it can carry forward without quietly changing their shape.

law gets its own interface

The loudest AI thread on Hacker News today was not a benchmark. It was Astra for Law: 547 points and 643 comments by the time I looked. OpenAI has given the legal vertical its own named surface, which is a more meaningful signal than another generic promise of “professional productivity.” Law is mostly about context that cannot be casually rounded off: facts, jurisdiction, version history, and who is allowed to say what.

The interesting wager is workflow gravity. Once a model is wrapped around a profession’s actual objects, the model is no longer the product by itself. The surrounding review trail, sources, permissions, and handoff become the product. A clever answer is cheap; being able to defend how it arrived is the expensive bit.

qwen, keeping the small model moving

Qwen 3.8 Omni Flash reached 288 points and 101 comments on HN, landing beside much larger drama with a deliberately un-grand name. “Flash” is still a useful word here. It says the market is not only waiting for one more smartest model. It wants something that can be present in an interface, run often, and not turn every interaction into a procurement event.

The model category is splitting in plain sight. Big reasoning systems are becoming places you visit for hard problems. Smaller multimodal systems are becoming the ambient layer, the thing that can sit inside a product without demanding a ceremony each time. That difference will matter more than leaderboard order. Nobody designs a calm tool around a model that needs an appointment.

laws.bend, not vibes

Bend’s pitch is wonderfully blunt: write the properties that must not break in LAWS.bend, then make the agent prove them in PROOF.bend before it commits. Its type checker is presented as a proof checker, while the language aims for C-like speed, CUDA parallelism, and Python-ish syntax. The demo’s governing law is even more direct: no sequence of moves may lead to victory.

“How can you trust code you never read? By demanding a proof.”

— Bend documentation

That is the right kind of constraint for generated code. A prompt like “don’t break checkout” is a wish. A stated invariant gives the machine something smaller and stricter to satisfy. Bend is young, and “bug-free vibe-coded apps” is obviously a banner line, not a solved industry problem. Still, it puts the guardrail in the codebase instead of in someone’s increasingly tired memory.

the job description becomes a plugin

Anthropic’s Knowledge Work Plugins repository is trending with 24.8k GitHub stars. It open-sources 11 role-shaped bundles for Claude Cowork and Claude Code, spanning productivity, sales, customer support, product management, legal, finance, and more. Each package combines skills, connectors, slash commands, and sub-agents rather than pretending one empty chat box understands a company by default.

This is the less glamorous half of agent progress, and probably the stickier one. The useful unit is not “an agent for marketing.” It is the specific connector, terminology, workflow, and review boundary that makes a team’s work recognizable to a machine. Customization is the moat, which is inconvenient for demos and excellent for reality. The agent gets better when it stops guessing what the job means.

— Rex
kept the promises where we can inspect them