Monday's brief starts at the edge of the agent loop, where the web decides whether to let you in and every page you fetch gets billed again as tokens. Then two open releases that run on hardware you already have: NVIDIA's whole GPU scheduler, and a forecasting model from Mistral that doesn't need a GPU at all.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Yelp, Delta, JPMorgan and Schwab are blocking AI agents. Amazon is turning away Meta's Muse.
The Verge pulled together a month of reports on October 17. Yelp told TechCrunch it only lets in agents that pay for its data licensing program. Delta is stopping agents from booking tickets for users. JPMorgan Chase and Charles Schwab confirmed they block agents from customer accounts over security concerns. Amazon said Meta's Muse's continued access violates its policies, and users of xAI's Grok, OpenAI's Dots and Instinct report the same wall there. Users say Cars.com and Amtrak stopped working too, which suggests more companies are blocking quietly.
Meta's answer has been deals. Walmart, Gap, Sephora and Best Buy let Muse shop, and Shopify opened its storefront network to agents. Explicit bans aren't the whole story, though. EquiStamp CEO Chris Canal points out that ordinary anti-bot systems, built for scammers, trip on agents just as easily. Reddit posts claim Google has flagged some accounts that used Muse for "suspicious activity." Those are user reports, and Google didn't answer The Verge's request for comment, but they line up with everything else in the piece.
For builders
If your product drives a browser on someone's behalf, log every 403, CAPTCHA page and login challenge by domain, and build the handoff now: when a site refuses the agent, hand the user a prefilled link or cart and stop retrying. Retries from a user's own logged-in account are how that account gets flagged. For commerce, check whether the merchant is on Shopify's agent catalog before you scrape its site.
Story II
TinyFish made search free. Its own fetch test shows why the page size still matters more.
TinyFish's Search and Fetch now cost nothing within limits: Search at 30 requests a minute and 500 an hour, Fetch at 150 URLs a minute and 1,000 a day, and both keep working on a $0 wallet. Its Agent product is $0.016 a step and Browser is $0.002 a minute. On Artificial Analysis's September 28 search leaderboard, which fixes the answer model (GPT-5.6 Luna, medium) and swaps only the search provider, TinyFish scored 71 at $9.30 per 1,000 tasks, the cheapest total of the 12 configurations shown. It was also the second slowest at 59 seconds a task. Perplexity Search (medium) topped the board at 80. No search at all scored 33.
The more useful part is a fetch test TinyFish ran on itself across 45 URLs and five providers. For one Daily Mail article with about 4,300 characters of body text, TinyFish returned 4,737 characters, Tavily 46,433, Exa 65,793 and Firecrawl 171,095. MarkTechPost estimates the Firecrawl result at roughly 42,774 tokens, a third of a 128K window before the agent has said anything. That's one page, run by a vendor on its own product, so don't generalize the ranking. The point holds anyway. Free retrieval saves you the request fee, and every agent turn resends whatever you fetched.
For builders
Add one line of logging to your fetch tool today: characters returned per URL, and the share that survives your cleanup step. Then set a hard cap (8,000 to 12,000 characters is a reasonable start) and strip nav, cookie banners and related-article blocks before anything enters context. Rerun your eval with the cap on. If quality holds, you've cut input cost on every turn of every loop.
Story III
NVIDIA open-sourced all of Run:ai under Apache 2.0, and you can run it on a laptop with fake GPUs
NVIDIA released the full Run:ai GPU orchestration platform on October 13 as a monorepo at dsx-ai-factory/run-ai: control plane, cluster services, installer, CLI and an MCP server for AI assistants. It runs on top of Kubernetes or OpenShift and ships as four pieces you can adopt separately. KAI Scheduler, already a CNCF Sandbox project, handles gang scheduling, fair-share and bin-packing. A GPU Fractioning Operator uses CUDA 13.4 so the driver enforces a hard memory limit per workload. Resource Management adds projects, quotas and node pools. Workload Map is a Go library that describes job types like PyTorchJob and KServe.
The local setup is one script. ./local-dev.sh builds about 50 Go services in containers and stands up a five-node Kind cluster on Kubernetes v1.32.5, with a fake operator simulating eight Tesla A10s. On Linux it wants 4 CPUs, 16 GB of RAM and 100 GB of disk, and a full build leaves around 44 GB of caches. The free community edition has no NVIDIA support. NVIDIA published no platform benchmark, though Microsoft measured the related Model Streamer loading models 3x to 6x faster on eight H100 VMs.
For builders
If you serve models with vLLM or SGLang, the cheapest thing to try here is the streamer: add --load-format runai_streamer and time a cold start against your current loader. If you share GPUs across a small team, read the fractioning operator's docs before buying another card. Splitting one GPU between a notebook, a small model and a dev session with enforced memory limits is now free software.
Story IV
Mistral's Timestral forecasts time series on a CPU, and its weights allow commercial use
Mistral released Timestral, its first time series family: four dense models at 5M, 28M, 123M and 209M parameters. The largest is 837 MB in fp32, takes up to 32,768 time steps of context, forecasts 1,024 steps in one pass and returns 21 quantiles. It handles several series at once with past-only, known-future and categorical covariates, which plain ARIMA fitting one series at a time doesn't. Mistral says the 209M model ranks first among zero-shot pretrained models on GIFT-Eval, TIME and fev-bench. On GIFT-Eval its MASE is 0.658 against TimesFM-3's 0.667. That's a narrow lead, and it's Mistral's number.
The license is the bigger difference. Timestral's weights are under a modified MIT license and the code is Apache 2.0, while TimesFM-3's weights are non-commercial. Mistral lists the 209M model as a public preview.
For builders
Got a metric you currently forecast with a spreadsheet trendline (signups, API spend, queue depth)? Run pip install git+https://github.com/mistralai/timestral.git, load TimestralForecaster.from_pretrained("mistralai/Timestral-209M"), and call forecast(context, prediction_length=30) on a (series, variates, time) array. Backtest it against the last 30 days you already know, and use the 0.9 quantile, not the median, for capacity planning. Read the modified MIT terms before you ship it.
— The Vibe Gate news desk. We read the firehose so you can keep building.