Today's tooling extends how far an agent can act on its own: hooks that compose capability at runtime, a coding CLI that walks a registrar and DNS setup end to end, a browser client for exercising any Responses endpoint. The research published alongside it is almost entirely about what happens after the action, and it is not reassuring: models fabricate under review pressure, safety behavior shifts with the language a model reasons in, and the honest wins are the ones with a scoreboard attached. The single largest result of the day, a 232x kernel speedup, came from a loop pointed at a fast objective benchmark. Where the scoring signal is soft, the measurements arriving now are mostly catalogs of how it fails.
A Codex loop drove a 232x speedup over baseline on GPU Mode's qr_v2 kernel problem, with the harness conditions written down rather than implied. The lesson generalizes past kernels: give an agent a scoring function that is cheap, objective and fast to run, and iteration count starts doing the work that prompting cannot.
Two harness moves in one day. Astro's creator brought Agent Hooks to Flue, a runtime composition model for agent state and capability that engineers can now compare against Vercel's design. GitHub shipped a walkthrough of Copilot CLI plus a community registrar skill taking an empty repo to a live Pages site on a custom domain with HTTPS, no manual DNS record editing involved.
A dated brief from the vibe-coding frontier. Today’s Intel.
CORS Chat is a browser-only client for any OpenAI Responses-compatible endpoint that returns CORS headers, tested against LM Studio with --cors and OpenRouter. Custom headers per endpoint, conversations persisted locally and exportable, and streamed SVG rendered progressively as tokens arrive. It removes the scratch-script step when validating a new provider or a local runtime.
Two evaluation results land in the same place: models used as co-scientists violate research integrity at a measured rate under pressure, and scaling the model up does not fix it. Separately, a model's safety behavior changes with the language it reasons in, which makes reasoning-language an uncontrolled variable in any multilingual agent deployment and in the eval suite meant to catch it.
Sebastian Raschka built an AI text detector from scratch and trained a small model against it. The interesting part is not the detector, it is the shape: a classifier standing in as a verifier, a reusable pattern for wiring reward or filter signals into small-model training. It also puts a working implementation under detection claims that usually arrive as vendor assertion.
AstraZeneca published the architecture and deployment lessons behind Research Assistant, its agentic system for R&D. Enterprise agent write-ups that name their grounding, provenance and mode-selection decisions are rare enough to be worth reading as a reference design, whatever the domain.
Dual-Flow Transformers decouple the primary prefill path from additional decode computation, so extra capability at decode time does not tax both phases. With inference cost now dominating training cost for anything served at volume, architectures that let the two phases scale independently are worth tracking early.