17 stories, 70 Manifests cited.
Three preprints reframe retrieval as an agent skill rather than a lookup: SpectRAG selects context under a token budget, a metacognitive supervisor decides when search is needed, and GraphNIAH shows frontier models quitting multi-hop searches too early.
Four preprints map the new security stack for agentic AI: AgentShield defends web apps, learned attention gates harden models against prompt injection, ANGLER stress-tests web agents with an adaptive attacker, and MIRAGE audits bias where agents actually decide.
Four preprints attack the same bottleneck from different sides: slim attention halves context memory, smaller compressors beat bigger ones, a Muon variant drops momentum buffers, and a quantization method shrinks vision-language-action models without losing policy performance.
Sakana AI is moving into pretraining its own models, DeepSeek Harness is sketching a Super-agent runtime, and a preprint describes orchestrators that improve through self-play. Together they point to agents built from the ground up rather than bolted onto existing models.
Serval's blogs describe service management, MCP connections, and chatbots converging on agents that resolve requests end to end, while Presight's CEO warns most companies still are not getting material value from AI.
Modal's CEO argues raw token-selling is a race to the bottom while the market splits by latency and workload, and new work from FriendliAI, fal, and a chiplet architecture paper shows what specialization looks like in practice.
A human-annotated benchmark finds over-proactivity is the dominant failure mode of proactive agents, while new work on change-blindness, approval, and verification shows what a more careful agent would do instead.
Five new papers push robot and game models from perceiving a scene to forecasting it at several horizons, then using the forecast to pick actions.
Four ChemRxiv preprints show agents moving into the chemistry workflow: LiteratureAgent turns papers into databases, XinPilot traces high-throughput synthesis, ETALON verifies agentic drug design, and vibe-qc keeps expert scientists in charge of agentic coding.
Macquarie and Wellington open late-stage venture to wholesale investors, PRENEU raises a Series D, and transport sees big deals alongside regulatory pressure.
Deere, Vertical Aerospace, Ur-Energy and Advantage Solutions all filed leadership changes within days, from board seats to severance terms.
Tempus, Marriott and Axsome reported growth and raised their outlooks, while Loews and CNA counted the cost of long-term care reserve updates.
New Zealand surprised markets by standing pat, Japan held while edging toward its inflation target, and Britain looks set to hike, while Slovakia warns patience has limits.
Four bioRxiv preprints show models moving into the wet-lab workflow: GENESIS predicts single-cell responses near experimental reproducibility, ScanFold3 and LIMBER read RNA structure at scale, and FateDriver ranks the regulators that drive cellular reprogramming.
Olson and Dubón repeated in Atlanta, Bellinger won again in left field, McLean ended the Mets drought, and Walls took shortstop honours in Tampa Bay.
Detroit ended Minnesota first loss with a ground game, Denver rallied past Kansas City, New England came back on Miami, and the Rams shut out the Chargers after halftime.
Britain FCA reviews its whistleblowing record and its PRA lifts a lending cap, while Singapore MAS cuts capital charges for infrastructure and backs insurers with a grant.