Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

Enhans raises KRW 51.2B for AgentOS
09/17/26, 11:46 AM

Enhans raised KRW 51.2 billion ($37 million) in Series C funding, taking cumulative financing to KRW 81 billion for its enterprise-agent platform. Seoul Economic Daily’s report says the round will fund AgentOS, while the company’s AgentOS documentation describes a stack that connects enterprise data, ontology, task agents, workflows and ACT-2 execution across existing systems. Operators should file this as another sign that enterprise-agent budgets are moving toward execution and governance layers—not chat wrappers—alongside the agent-evaluation economics already visible in benchmark work.
Wire

D-Robotics closes a $400M Series C
09/17/26, 11:44 AM

D-Robotics closed a $400 million Series C, earmarking the round for its Sunrise chip portfolio and a software stack for robot development. The company’s funding announcement says Sunrise shipments have passed 8 million units, while the S600 reached more than 20 embodied-AI customers within six months and first-half revenue grew several times year over year. For builders, this is a capital-and-distribution signal: compare any robot compute plan against the cost curve emerging in Unitree’s robot stack before assuming model quality is the bottleneck.
Wire

Salesforce harness lifts browser-agent scores to 93%
09/17/26, 10:48 AM

Salesforce’s DarwinX paper reports that evolving an agent harness—without changing model weights—raised WebArena-Infinity real-task pass@1 from 43.5% to 93.0% audit-clean after evolution on 300 synthetic intents, with 1,260 unseen real tasks in evaluation. The primary paper defines the harness as prompts, tools, skills, and control flow, making scaffolding—not only model choice—the tested lever. This remains a benchmark result, not production reliability: agent teams should budget for harness and evaluation work and reproduce it on private tasks, as the archive’s agent-evaluation cost analysis shows.
Wire

Firefox brings Mistral Small 4 to three markets
09/17/26, 10:47 AM

Mozilla is adding Mistral Small 4 to Firefox Smart Window beta across three launch markets—the US, Canada, and France—while keeping model choice open, according to Mozilla’s announcement. Mistral’s release says conversations are not saved on Mozilla’s servers by default and Mistral agrees to zero data retention. Browser-agent builders should read this as a distribution and privacy boundary, not a benchmark win: test provider switching, browser-context handling, and retention terms before embedding Smart Window-style flows, following Mistral’s sovereign-AI deployment economics.
Wire

Rune raises $40M for solar AI compute
09/17/26, 10:46 AM

Rune raised a $40 million Series A for RELIC, a modular system that puts AI compute at solar sites and brings reported total funding to $53.5 million, according to its launch release. The release says solar plants can waste up to 20% of generated energy and RELIC can bring GPUs online in as little as six weeks; secondary coverage describes the same stranded-power pitch. Operators should treat this as a site-specific alternative to grid queues—verify curtailment rights and utilization before counting solar capacity as dependable AI supply, as Google’s Finland power deal separates preserved power from new capacity.
Wire

Dream-RSI cuts discovery calls 162×
09/17/26, 09:48 AM

Dream-RSI reports up to 162× fewer agent calls than SimpleTES while maintaining or improving discovery quality across eight scientific tasks spanning algorithm engineering, mathematical optimization, and GPU kernels, according to the authors’ arXiv paper. It replays completed discovery trees to test exploration policies off-policy, so the headline is a search-cost result—not evidence that model weights recursively improve; on the Lasso task, the paper reports 1.7× fewer calls than fixed exploration. Builders extending recursive self-improvement research should measure cost per validated discovery, not call-count gains on one benchmark.
Wire

MIT adapts surgical scans in five minutes
09/17/26, 09:47 AM

MIT researchers’ xvr model adapts to a new patient in about five minutes, then aligns X-rays with 3D scans in seconds at sub-millimeter precision, according to MIT’s research account and the Nature paper. The foundation model was pretrained on whole-body scans from more than 2,000 patients and outperformed existing AI methods by an order of magnitude in the reported evaluation; it remains research validation, not a clinical deployment guarantee. Builders tracking AI’s clinical validation gate should file this as a promising registration layer, with prospective hospital evidence still the switch trigger.
Wire

GitHub ports Copilot runtime to 832,378 Rust lines
09/17/26, 09:46 AM

GitHub says Copilot’s agent runtime reached 832,378 production Rust lines plus 468,689 Rust unit-test lines by August 21, after agents wrote most of a port merged through 128 pull requests, the company’s engineering account says. The project moved incrementally and kept end-to-end tests running; GitHub’s claim is a case study in one vendor’s workflow, not a general license to rewrite every TypeScript system. Teams evaluating Copilot’s model-routing economics should copy the control loop—small merges, tests, rollback—not the language choice.
Wire

Odyssey-3 drives 77% as far as real-trained policy
09/17/26, 08:46 AM

Odyssey-3’s driving policy traveled about 77% as far between safety-driver interventions as a policy trained on real footage, after learning from only 20 hours of simulated driving data, the company reports in its launch account. The same world model is being tested across robot arms, humanoids, drones, vehicles, and games, but Odyssey says public release is still weeks away, making this a company-run transfer result rather than a deployment benchmark. For builders tracking the archive’s physical-AI funding thesis, implication is simple: world-model vendors are selling data efficiency, and repeatable cross-body evaluation is the proof point to demand.
Wire

Ferry Health launches with $9M care agent
09/17/26, 08:44 AM

Ferry Health launched from stealth with a $9 million seed round and says its care-navigation agent is already available to more than one million patients across all 50 states, according to the company’s launch announcement. Its agents identify in-network providers, call to verify availability and coverage, and schedule appointments; the planned fivefold expansion is still a forecast, not utilization evidence. Builders can file it beside the archive’s analysis of medication-access agents as a signal that health AI’s near-term bottleneck is phone-and-insurance workflow execution, not another diagnostic demo.