Debugging agents means querying traces, not scrolling logs
Braintrust, LangChain, Databricks and Andesite all point the same way: keep full traces, query them like data, and let humans judge what they show.
28 stories, 132 Manifests cited.
Braintrust, LangChain, Databricks and Andesite all point the same way: keep full traces, query them like data, and let humans judge what they show.
Five preprints test agents on cross-repo changes, self-caught bugs, continual learning, code review and performance work. The pattern is consistent: the flashy demo works, the maintenance work does not.
In four preprints, alignment methods that look solid in one setting weaken when values compete, images enter the loop, inputs are attacked, or safety is tested outside the training distribution.
An 80-gigawatt shortfall by 2029, permitting delays pushing revenue into 2027-2029, and a $60 million property deal bought for its megawatts.
Three papers point the same way: agents get better when they keep what they learn, pass reasoning to the right peer, and face guardrails that can be checked rather than trusted.
Six preprints map a widening attack surface: poisoned pipelines, compromised tool updates, persistent memory, typographic tricks, malicious skills and peer pressure inside agent populations.
SoundHound wants the living room, Resemble wants the meeting room, Apple is opening Siri to developers, and a voice dictation tool lands on Linux. The common thread is voice becoming the interface — and the fraud risk that follows it.
Keeping data inside a border is only one layer. The harder work is controlling the models, the compute and the access around them — and most governance teams have not covered all four.
In five preprints, long-term memory and retrieval systems break most often when they fetch evidence, when their knowledge graphs are noisy, and when their stores are split apart — and the fixes are checks, smaller memories and federated search.
Insurers, lenders and document teams all converge on the same discipline: the model proposes, a deterministic check or a human verifies, and the result can be traced back months later.
Fireworks' ecosystem, Glean's routing, Toloka and Preference Model's evaluation data, and Baseten's serving layer all point to a stack, not a download.
Corti, Invisible, Ambience and Rad AI all measure the same thing: visits, throughput and charting hours, not model scores.
Supabase, E2B, Intel and Nevermined all shipped pieces of the same puzzle this week: identities for workloads, permissions enforced outside the model, and payment channels for sub-cent calls.
In five preprints, single-cell methods reconstruct how cells choose fates over time, separate how many cells of a type are present from what state they are in, read chromosome structure from single cells, and turn microscopy into interpretable phenotypes.
In four preprints, cancer classifiers reach near-perfect scores by learning which dataset a sample came from, breast spectroscopy accuracy collapses once patient leakage is removed, and harder benchmarks expose how far AI research help still has to go.
Talen upsized its programme to $3.0 billion, Procore, Latham, Primoris, DuPont and Strata all announced or executed buybacks, and the scale ranges from $20 million to billions. The signal is not the same in each case.
UBS raised its AI spending forecasts to $1.0 trillion for 2026 and $1.4 trillion for 2027, Super Micro guided to as much as $82 billion of annual sales, and DigitalOcean is funding capacity with a $785 million lease facility. The build-out is being financed as if it has years left.
Fiserv took a $6.45 billion goodwill charge, Driven Brands is restating three years of accounts, Vital Farms withdrew its 2030 targets, and AdaptHealth and Solid Power reset guidance. The clean-up is the story.
The Banque de France wants strategic control of euro payment infrastructure, Iceland is building its own account-to-account scheme, SEPA still reaches fewer than 40% of surveyed consumers in south-eastern Europe, and private options from marketplaces to exchange listings keep spreading.
Marathon is selling two refineries to CVR, Energy Transfer is buying midstream assets and approving a new pipeline, and Ovintiv and Talos are buying drilling inventory. The deals point in different directions by design.
Warp, OpenAI, Daytona and Traversal all point at the same shift: the bottleneck has moved from writing code to running, reviewing and fixing it at agent scale.
In four preprints, feeding duration shows no link to two infant outcomes, mothers with ADHD intend to breastfeed but rarely sustain it, a maternal marker tracks large birth size better than the standard test, and diarrhoea care still reaches richer children first.
In five preprints, resistant cancers rewire chromatin, metabolism and stress responses to survive treatment — and in each case the rewiring itself creates a vulnerability that a second drug can exploit, at least in laboratory and animal models.
Hometown Remix jerseys, a Nashville charity patch, a decade-in-the-making rink in Reseda and a free-sandwich promotion in Buffalo all trade on the same thing: belonging to a place.
The Bengals play the Falcons in Madrid, the Packers open a pop-up shop in Munich, and the NFL unveils a professional flag football league launching in summer 2027.
Third-quarter filings show revenue growth at Pfizer, Amneal, Madrigal and Rhythm, while the White House says prescription drug prices fell 4.5% year over year. The two pictures are not reconciled in either set of documents.
In announcements published November 3, two defense contractors agreed to pay tens of millions to resolve federal claims, while a former IT contractor pleaded guilty to spying for Ethiopia.
In five preprints, Antarctic ice loss jumps when melt lakes fracture ice shelves, sea ice breaks into a different floe pattern past a forcing threshold, ice-face melting switches drivers at a measurable speed, and warm seas open a storm path that vanishes when the warmth is removed.