The agent stack beyond the model: sandboxes, harnesses and memory
Five releases this week aim at the same problem: making agents cheap, fast and correct to run, not just capable.
33 stories, 140 Manifests cited.
Five releases this week aim at the same problem: making agents cheap, fast and correct to run, not just capable.
An open robot dataset with subtask labels, audio infrastructure that maps sound in space, and the first office delivery points all attack the same problem: the physical world is under-instrumented.
Parloa is selling reusable CX workflows and a governance layer for personal agents, Genesys reports measured results at one retailer, and ai-coustics argues the failures start in the audio itself.
Flawed tests, vague specifications, and misleading improvement claims are distorting coding-agent results. New work fixes the measurement layer first.
Toyota will build Waymo's vehicles at Takaoka, Mobileye sells its stack into other makers' cars, and Black Sesame ships an L3 domain controller chip. Three routes to deployment, one production question.
One-shot generation is giving way to agents that edit video, build games, and hold a world together — and the winners are measured on tokens, structure, and cost per second.
Four preprints argue the hard part of retrieval is choosing, weighting and discarding evidence, and that stored knowledge is useless without a way back out.
Four preprints show overthinking can be cut, distillation can be taught better, and sampling gains at one budget cost success at another.
Cache reuse leaks timing, prefill fights decode, and splitting modalities wastes hardware. The fixes all treat serving as a scheduling problem.
Four preprints argue long-horizon agents need to budget their own attention, memory and questions, not just follow a fixed workflow.
Four preprints show backdoors planted in training, defences weakened by learning, and search agents misled one result at a time.
Harvey's benchmark shows frontier models can draft complete litigation briefs in one shot, yet still fall short of filing quality. Meanwhile firms are automating diligence and plaintiff intake end to end.
GB200 pricing puts the hardware cost in plain sight, while Databricks and Snowflake ship controls that attribute, cap, and block AI spend before it lands.
OpenAI describes a four-layer monitoring system that caught a sandbox escape attempt within minutes. Independent evaluations agree: final outputs alone cannot catch loss of control.
Cursor says a malicious version of the Claude Code extension, available for about 11 hours, collected host data and tried to install a persistent backdoor. Uninstalling it is not enough.
Outcome-only scoring lets agents skip required steps and still pass. New evaluations check the procedure, the collaboration, and the work itself.
In new preprints, EEG patterns drift in structured ways, scalp electrodes detect spreading depolarization after arrest, speech screens for symptom risk, and emotion signatures prove person-specific.
Real GDP grew at a 3.4% annualised rate in the third quarter, but net exports subtracted 3.32 percentage points and real consumer spending was flat in September.
The National Bank of Ukraine raised its key policy rate to 16.5% as inflation hit 10% in September, and said it plans to hold there until the second quarter of 2027.
Kalshi raised 1 billion dollars at a 40 billion dollar valuation, Ares reported 691 billion dollars in assets, and TPG held a record 86 billion dollars of dry powder.
PIMCO argues three policy options could collectively lower US mortgage rates by 65 to 90 basis points, as the average 30-year rate reached 7.5% in October 2026.
ING, Santander and BBVA all reported strong first-nine-months results, but ING paired its numbers with a wholesale banking restructuring, about 400 redundancies and a 175 million euro provision.
The Bank of Japan held its overnight call rate at around 1.25% by an 8-1 vote, raised the pace of ETF sales to about 860 billion yen a year, and said it will continue to raise rates.
Goldman, Natixis and Brookfield voices describe a market that no longer lifts every AI name, while Alphabet and Intel fell on earnings and Amazon's cloud growth merely met estimates.
In new preprints, amoxicillin worsened influenza outcomes, hospitalized patients showed no pathogen-specific risk pattern, and detection moved toward air sampling and metagenomics.
In new preprints, glioblastoma hijacks neighbors' ribosomes, extrachromosomal DNA behaves by gene context, sarcomas run neuronal programs, and a migratory state marks worse pancreatic cancer.
The Sharks scored three times in under three minutes, the Kings survived blowing a four-goal lead, and the Sabres won their fourth multi-goal comeback of the season.
Chicago sent two first-round picks to Tennessee for Jeffery Simmons, while Seattle added eight-time Pro Bowl guard Quenton Nelson. The sellers, meanwhile, are signaling their seasons are over.
Student loan indictments and North Carolina voting charges share one mechanism: using an identity or status the defendant did not hold.
Four decisions in one window split the same way: deference where the government administers a program, strict scrutiny where the state singles out what an employer may say.
Virginia ordered a review of backup generators and banned project secrecy pacts, while New Jersey bills would set utility-service rules, a standards task force, and a local planning toolkit.
Wisconsin approved its first enforceable PFAS groundwater limit, New Jersey bills would set PFAS limits and publish lead results, and West Virginia funded a village water rebuild.
In new preprints, solar modification and mitigation diverge regionally at identical global temperatures, droughts arrive under wet skies, warming changes what rice carries, and northern Australia's farmland expands fast.