11 articles — 🟡 3 important , 🟢 8 interesting
The vLLM Semantic Router has gained Session-Aware Agentic Routing (SAAR), a mechanism that understands long-running agentic conversations instead of treating each message separately. Across 21,600 test turns, SAAR cuts model switches by 79.29 percent, fully eliminates unsafe switches, and lowers estimated costs by 78.71 percent.
vLLM has integrated Intel's AutoRound quantization into vLLM-Omni, enabling W4A16 compression of multimodal and diffusion models. The result is up to a 62 percent smaller checkpoint, with minimal quality loss and faster generation on Intel XPU and NVIDIA graphics cards.
OpenAI has introduced a new set of capabilities for Codex that extend the tool beyond engineering tasks toward broader business roles. The goal is for analysts, marketers, designers, investors, and other teams to be able to use Codex in their everyday workflows.
The new Curation-Bench benchmark tests whether coding agents can automate the curation of training data. Agents reach strong baselines within ten iterations, but exploring diverse policies requires structured scaffolding that pushes them to adapt existing methods.
On June 2, Anthropic introduced two changes to the Claude API. The advisor tool now supports the max_tokens parameter to cap the advisor model's output per call, and requests that return stop_reason refusal with no generated output are no longer billed.
LangChain Labs, in collaboration with Harvey, has published research on cheaper verification of legal AI agent outputs. Batching verifiers and using open models cuts the cost per criterion by roughly ten times, while keeping quality comparable to frontier models.
Anthropic has expanded Project Glasswing, a program to secure critical software infrastructure, from an initial fifty to roughly 150 new organizations across more than 15 countries. So far partners have found over 10,000 high- and critical-severity vulnerabilities with the help of the Claude Mythos Preview model.
Researchers have presented a framework for vetting AI agents before they are deployed in regulated industries. The system uses an ontology to generate scenarios and issues machine-verifiable trust certificates, outperforming persona-based approaches in regulatory coverage.
