The Scientist

TheSequence reports that RRSI, a harness-search method from Google Cloud AI Research, Stanford and others, gained +3.4 points on six held-out benchmarks with a frozen Claude Opus 4.8. Any loop that edits your prompts or tools against a fixed eval set ends up fitting that set, so its wins need held-out proof. Alejandro Saucedo's account puts the gain at +3.9 (43.6 vs 39.7), so check the paper before you quote either figure.

TheSequence reports that NVIDIA and KAIST's Mid-Harness lifted the TMAX-9B agent's Pass@1 from 50.00% to 68.03% by having GPT-5.6 Sol check candidate actions before they run. Separately, Microsoft and KAIST's CorpusMap added +6.45 RAG quality while using 0.43–0.65× the input tokens. You can try both around your current models without retraining anything. The cheaper self-verification gains, though, come to about two tasks and sit inside the noise.

The Information reports that Nscale is targeting a $35B IPO valuation on the strength of a $103B contracted backlog. Nearly all of that backlog depends on 12 data centers that are not yet built or fully financed. Nscale's Anthropic, Microsoft and Figure AI deals ramp in 2027–2028, so part of your future Claude and Azure capacity depends on that construction. The Financial Times reports that Nvidia is in early talks to insure lenders against losses on GPU collateral at smaller clouds.

Alejandro Saucedo's write-up reports that Uber's MCP platform catalogs 800+ servers and 5,000 tools, and that its scaling limit was context bloat. Uber cut the bloat by discovering tools on demand, letting agents request only the response fields they need, and writing tool outputs to files. It also registers every tool as disabled by default. In your own agents, the tokens you spend on tool schemas each turn trade directly against tool-selection accuracy. OpenAI's always-on Dots agents, with 4,000+ app connectors, widen the same permissions surface.