Stay ahead in AI without the scroll.

10 minutes a day: what changed, why it matters, and what to try next. Free, every morning, 7 days a week.

A steaming ceramic tea bowl painted with blue leaves, a small blue ceramic whale and a brass key, on a weathered wooden windowsill, with a misty harbour and sunrise over the hills outside

· 15 min read

DeepSeek opens its app tools via API; Anthropic maps always-on agents

Published

Since 13 November, the tools behind the DeepSeek App are callable through the DeepSeek API, and DeepSeek open-sourced a static chat client. The same weekend, Anthropic published a blueprint for always-on agents built around an inbox, timers, memory, permissions and an outbox.

Today in 30 seconds

  • DeepSeek now exposes the DeepSeek App's search, file, URL, safety and voice tools through its API, and open-sourced a static chat-demo client.
  • Anthropic published a blueprint for always-on agents: inbox, timers, memory, permissions and a quiet-by-default outbox.
  • Bloomberg reports banks are turning away China and Hong Kong orders for Anthropic's IPO, which still prices on 18 November as ANT.
  • LangChain raised $300M at a $3B valuation, pitching LangSmith as the agent control plane.
  • Sierra's Poppy 1.0 is live, and Shopify now publishes a Poppy file for every merchant automatically.

It’s a slow Sunday, and the weekend was mostly about plumbing. DeepSeek opened up everything behind its app as plain API calls, and Anthropic published a practical blueprint for agents that “never sleep”.

The bills for agent infrastructure kept arriving: LangChain at $3B, and Sierra’s Poppy protocol going live on every Shopify store. Also on the table: a 145B model that runs on a laptop, a benchmark of benchmarks that says most of them are already used up, and a Starship caught by its own tower.

Top news

Models & APIs

DeepSeek puts the whole DeepSeek App behind its API, plus an open-source chat client

Since 13 November, the web search, file reading, URL fetching, safety checks and voice features behind the DeepSeek App are callable through the DeepSeek API. DeepSeek also open-sourced chat-demo, a static web client that can rebuild the app on your own key.

  • The built-in tools run on V4.1-Flash (deepseek-flash) in the Messages, Responses and Chat Completions APIs. DeepSeek handles tool calls and context assembly on its servers.
  • What you can switch on:
    • Web Search (web or images, with citations, non-thinking mode)
    • Deep Thinking Search (searching and reading pages during multi-step reasoning)
    • File Input (PDFs, Office files, images)
    • URL Reading (links in user messages are fetched automatically)
    • Safety Checks (inputs and outputs screened, with refusal indicators)
  • The parameters are deepseek_web_search and deepseek_read_url. Search, Fetch, OCR, ASR and TTS are also sold as standalone APIs.
  • Rules: public-facing apps must enable safety checks. Developers stay responsible for moderation, model registration and labelling AI-generated content.
    • Tool use is billed on top of normal token prices.
  • deepseek-ai/chat-demo is a static browser client with no backend. It stores your key in localStorage and conversations in IndexedDB, and it can import your DeepSeek conversation history.

Why it matters: Search, files, URLs and voice used to be the glue every team rebuilt around a model. When the lab hosts that glue, a cheap model becomes a finished product stack, and it’s also lock-in: your agent loop now depends on DeepSeek’s servers, not just its weights.

Sources: DeepSeek API news, 13 Nov · Build a chatbot guide · github.com/deepseek-ai/chat-demo

Agents

Anthropic’s blueprint for “always-on” agents: inbox, timers, memory, permissions, outbox

A new Claude Platform guide (William Hu, 13 November) explains how to build agents that run indefinitely. The answer isn’t a bigger loop: it’s five durable pieces around the usual agent loop.

  • The key idea: an always-on agent is mostly asleep. It wakes up for a run, handles an event, saves state, and goes back to sleep.
  • Event inbox: webhooks, schedules and background jobs all write to one durable table, deduplicated by source IDs.
    • Only one worker per inbox runs at a time. The worker batches pending events.
    • The inbox stays the source of truth, so a crashed run can be retried.
  • Timers: follow-ups (deadlines the agent sets itself) and schedules (fixed cadences) live in their own store.
    • A scheduler polls that store and drops due timers into the inbox. Timers never start runs directly.
  • Memory: a directory of files on durable storage. It holds user instructions, user notes, past messages so nothing is repeated, and a progress file per task.
    • Rule of thumb: store only what the agent can’t look up again, and keep pointers for everything else. “Dreaming” can curate memory between sessions.
  • Permissions: a credential proxy outside the agent adds secrets to outgoing requests. Auto mode classifiers review risky actions.
    • Deferred tool calls end the run, save the session mid-call, show the user a card, and resume when the answer comes back as an inbox event.
  • Outbox: each run ends in one of three ways: send, stay silent, or ask.
    • Anthropic’s advice is to have agents “start quiet and then earn more of the user’s attention over time”.
  • Cost context from the post, on one long-horizon coding benchmark:
    • Opus 5.5 cost 78% less per solved task than Opus 5
    • Against 40% less on typical workloads

Why it matters: This is the first clear vendor reference architecture for proactive agents, and it lines up closely with how a multi-bot fleet with routines, a shared brain and approval cards already works. It’s a handy checklist for finding the missing piece: dedup, a single worker, or a quiet-by-default outbox.

Sources: Building always-on agents, Claude · @omarsar0

Markets

Anthropic’s IPO turns away orders from mainland China and Hong Kong

Bloomberg reports that banks on Anthropic’s IPO are refusing orders from institutional investors in mainland China and Hong Kong, partly because of US ITAR rules. The deal still prices on 18 November and starts trading on 19 November as ANT.

  • Several banks have refused such orders or think allocations are unlikely.
    • Communications with the New York syndicate reportedly cited ITAR restrictions. The same playbook was used for SpaceX’s IPO.
  • Investors are trying workarounds: a Chinese tech founder’s family office is reportedly ordering through a BVI trust, and others are looking at SPVs and swaps.
  • The prospectus lists no China/HK selling restrictions. It offers shares in the US, Australia, Canada, parts of the EEA, Japan, Switzerland and the UK.
  • Demand: reportedly about 6× the shares on offer.
    • Together with its co-founders, Anthropic is raising up to $90.4B (IPO plus private placement) at about a $1.4T valuation.
  • Goldman Sachs and Morgan Stanley lead the deal, with 20 other banks, and it lists on the Nasdaq Global Select Market.
  • Context: Anthropic’s products aren’t available in China or Hong Kong, and it is still fighting the Pentagon’s supply-chain-risk designation in court.

Why it matters: Frontier labs are now treated like defense-adjacent assets. That shapes who can own them, and over time who sets their governance.

Sources: Economic Times / Bloomberg · @AnikaCodes

Funding

LangChain raises $300M at a $3B valuation for LangSmith as the “agent control plane”

LangChain closed a $300M Series C (13 November) led by CapitalG. It is positioning LangSmith as the framework-agnostic layer for running, improving and governing agents.

  • New investors include GIC and DFJ Growth, plus Salesforce Ventures, Insight, QuantumLight and MongoDB Ventures.
    • Existing backers IVP, Sapphire, Sequoia, Amplify and Benchmark joined too.
  • More than 7,000 organizations use LangSmith. Named examples: Harvey, Rippling, NVIDIA, Toyota and Salesforce.
  • Clay monitors “hundreds of millions” of agent runs a month.
  • Workday cut the time to deploy a new agent from five months to under a day.
  • Product pillars:
    • Runtime: Deep Agents, LangGraph, Agent Server, Sandboxes, Managed Agents, Fleet
    • Improvement: Observability, Evaluation, LangSmith Engine fixing prompts and code, Fine-Tuning from traces
    • Governance: LLM and Tool Gateways for budgets, fallbacks and tool auth
  • More LangSmith features are due at a release event on Friday, 20 November.

Why it matters: The money is moving from frameworks to operations: evals, gateways and governance. “Own your intelligence” pitches against lab-hosted stacks, such as DeepSeek’s new API tools.

Sources: LangChain blog · TechCrunch · CapitalG

Agent commerce

Sierra’s Poppy 1.0 goes live, with every Shopify store and Meta’s Muse on board

Sierra’s Personal Agent Protocol (“Poppy”) reached v1.0 on 13 November, with Shopify now publishing a Poppy file for every merchant automatically. It defines how your personal agent talks to a company’s agents, websites and APIs.

  • Muse, Meta’s agent, now looks for Poppy first on every site it visits.
  • More than 100 design partners have joined, 60 of them new, including Delta, Hyatt, Intuit, Checkout.com, Cisco, GoodRx and Browserbase.
  • Mercury, Synchrony and Turkish Airlines are live.
  • It is built on MCP, OpenAPI and UCP, with OAuth so customers control what an agent may do.
  • Version 1.0 adds staged adoption, ways for companies to guide agents and collect feedback, and extensions, starting with a commerce extension co-written by Shopify.
  • How the pieces split: UCP does the shopping (cart, checkout), while Poppy handles the arrival (who the agent is and what it’s allowed to do).
  • The commerce extension lets an agent shop as a guest or signed in, and makes it ask the customer before sensitive actions.
  • Examples: Mercury customers can check balances and start transfers that still pass the account’s approval policy.
    • Turkish Airlines handles fares, bookings and business-class upgrades, which the customer approves.

Why it matters: Personal agents now have a handshake standard, not just scraping. If you run a site, a “Poppy file” may soon matter as much as a sitemap.

Sources: Sierra: Releasing Poppy 1.0 · The Next Web

Open models

Edge0-145B: a 145B open model built to run on a laptop

Edge0 released Edge0-145B and Edge0-120B (12 November) with its own inference engine. It reports these figures on a compatible laptop such as a MacBook Pro:

  • Peak memory: 29 GB
  • Speed: 80+ tokens/s
  • The design was co-built for the device from the start, not compressed afterwards. Each group of 64 experts shares two large matrices, and each expert keeps small parameters of its own.
  • Three storage tiers:
    • Hot experts stay in GPU memory
    • An expert library of about 40 GB sits in system RAM
    • About 48 GB of vocabulary data lives on the SSD
  • Loading overlaps with compute, plus speculative decoding.
  • Requirements from the model card: Linux or Windows with an NVIDIA GPU (16 GiB+ VRAM, 20 GiB RAM, CUDA 13), or Apple Silicon with the Q3_K build.
    • That needs about 98 GiB of SSD.
  • Reported results (Edge0’s own internal evals):
    • 64.6 on DeepSWE 1.1
    • A mean of 72.8 across six benchmarks
    • Some baselines are old (Claude Opus 4.6), and some competitor memory figures are estimates.
  • Edge0 says the full stack, including eval setup and reproduction steps, is being released.

Why it matters: “Frontier-ish on your own machine” moves from 30B to over 100B. Treat the benchmarks as claims until someone runs them independently.

Sources: Edge0 blog · HF: Edge0-145B-q3k · github.com/Edge0-AI/Edge0

Evals

SatBench: 63% of the benchmarks labs cite are “solved”, but only 24% by independent runs

Ethan Mollick published SatBench (14 November), a “benchmark of benchmarks” built on Brian Zhou’s Benchmark Almanac. It counts a benchmark as solved once the best reported score reaches 85%.

  • 109 of 173 benchmarks cited by labs since 2023 are solved (63%). The index sat near 30% until early 2025, then rose 20 points in 12 months.
  • New benchmarks fall faster each year. The share solved within a year of release went from 17% (2023) to 30% (2024) to 43% (2025).
    • The median time to 85% dropped from about 20 months to about 14.
  • The big caveat: counting only independent runs, SatBench is at 24%.
    • Three in four first crossings past 85% come from a lab’s own launch post, and only 10 of 75 were later matched independently.
  • The 2023 canon is gone. In 2026, 2 of 134 launch posts named MMLU, HumanEval or GSM8K-style classics.
  • First past 85%: OpenAI on the most benchmarks overall (20), Anthropic in 2026 (11).
  • Extrapolation, not a forecast: today’s set would be 85% solved around 2028, and the live index around 2030.

Why it matters: Launch-post scores no longer tell frontier models apart. Unsolved benchmarks (HLE, SciCode, IFBench) and your own real-work evals do.

Sources: github.com/emollick/satbench · @emollick

Evals

SUS-Bench: 28 models played a Mafia-style game with 250,000 humans, and a fruit-fly brain beat 17 of them

People Make Things published SUS-Bench (13 November). It ranks models by whether their team wins more often in a live social-deduction game with real people who don’t know who is human.

  • Scale, as reported by the authors:
    • 250,000+ players from 167 countries
    • 500,000+ rounds, 1M+ meetings, 90M AI decisions
    • Players weren’t paid.
  • Top of the board:
    • Claude Fable 5.2 (51.0)
    • GPT-6 Astra (47.9)
    • Claude Fable 5.1 (45.9)
    • GPT-6.1 Terra (44.0)
    • GPT-6.1 Sol (43.2)
  • Seven models beat the human baseline (about 35).
  • A simulated fruit-fly brain steering one player scored 13.0 and finished above 17 of the 28 models, including Grok 4.7 (7.9) and DeepSeek V4 Pro (7.3).
  • Claude Haiku 5.5 came last (2.1).
  • Behaviour clips: GPT-6 Astra turns a witness’s story against her (57.5 imposter wins per 100).
    • Claude Opus 5.5 votes out a cornered partner 100 times out of 100, while Grok gives its partner alibis.
  • Cost angle:
    • GPT-6.1 Terra reaches 44 points at about $0.22 per game
    • Against $4.41 for Fable 5.2’s 51

Why it matters: It’s a fun, noisy benchmark, but a rare one: models are scored on cooperation and deception with real humans, not on static questions. Read it as an experiment, not a ranking to quote.

Sources: @imodarressi thread · peoplemakethings.com/sus · vc.ru write-up

Agent infra

Cloudflare open-sources Kitesurf, its container-free browser for agents

Kitesurf is now on GitHub under Apache 2.0 (12 November). It’s a browser engine that runs entirely on Workers: it parses HTML, applies CSS, runs page JavaScript and renders screenshots, with no Chromium container.

  • It speaks the Chrome DevTools Protocol, so Puppeteer and similar CDP clients connect directly. WebMCP support arrived in September.
  • It’s built from Rust crates compiled to WebAssembly. It has three parts: Engine (sessions and clients), PageScript (the page and its JS) and PageRenderer.
    • JavaScript runs in Dynamic Workers when possible, with an in-WASM engine only as a last resort.
  • You can run it locally with wrangler or deploy your own instance. Self-hosting needs a Workers Paid plan.
  • Cloudflare says it’s under heavy development, with compatibility as the focus.

Why it matters: Headless Chrome is the heaviest part of most browsing agents. A lightweight CDP-compatible engine you can self-host cuts the cost per page for scraping and research agents.

Sources: Cloudflare blog · github.com/cloudflare/kitesurf · Changelog

Policy

SpaceXAI reverses course: Grok API text watermarking is now off by default

A week after Grok API text watermarking was supposed to become default-on (from 7 November), SpaceXAI’s docs now say text watermarking is off by default. Elon Musk confirmed it on 14 November: “only on where required by law or requested by the user”.

  • Team admins can turn it on per model in the console. Changes apply within about 5 minutes, and every opt-in and opt-out is logged.
    • Some models may be locked on for a team.
  • Audio from the voice APIs stays watermarked by default. You can opt a voice model out, except where the user’s location requires the watermark.
  • Turning watermarking off means confirming that a legal exception applies, and that you’re solely responsible for any marking or disclosure.
  • It’s invisible and token-level (SynthID Text), with no characters added.
  • An x-watermark-style response header tells you if a reply was marked, and detection is statistical, with access granted on application.

Why it matters: EU AI Act marking duties now fall on app builders, not the model provider’s defaults. If you ship Grok text to EU users, decide on purpose whether to switch it on.

Sources: SpaceXSI docs: Output Watermarking · The New Stack · @elonmusk

GitHub repos

  • elementalsouls/Claude-BugHunter

    An offensive-security skill bundle for Claude Code: 83 skills and 15 slash commands, with hunting patterns drawn from 681 disclosed HackerOne reports across 24 vulnerability classes. It also covers enterprise perimeter chains (M365/Entra, Okta, vCenter, SSL-VPN, cloud IAM), triage and reporting, and a “7-Question Gate” before you file. It installs as a plugin (/plugin marketplace add elementalsouls/Claude-BugHunter). Use it only on targets you’re authorised to test.

    MIT, ~5.9k★, +342 this week

    View repo →
  • Tencent/LoopForge

    A self-hosted workspace for shipping with coding agents. A “Main Bot” opens a Docker sandbox, checks out your repos, delegates to project roles and opens the PR. Repeatable SOPs chain Agent, Script and Human-gate nodes. It drives CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Kimi Code, Pi and others.

    created Jul 2026, ~1k★

    View repo →

Research papers

  • TOME-Math

    Turns 2,118 university math textbooks into about 380K structured proof-supervision records, keeping chapter order and proposition-proof links. Fine-tuning on it lifts mean scores by 16.7 points.

    2610.40394

    Read paper →
  • SparseRead

    A reading-control layer for web agents that admits only the evidence an agent needs. It cuts total tokens by up to 92.9% and wall time by up to 89% across six models, with accuracy preserved or improved.

    2608.22237

    Read paper →
  • Ropedia Xperience 2

    1,589 metric-scale 3D scans plus thousands of egocentric recordings, built to study how physical-AI capability scales with environments and behaviours.

    2611.17060

    Read paper →
  • WorldDraft

    Asks: if code builds the 4D world, can video models just render it? It tests that with 432 paired cases across 48 executable world families.

    2611.16217

    Read paper →
  • FOAM

    A lightweight SOAP variant with input-only preconditioning, aiming for second-order optimizer gains at close to AdamW’s memory and speed.

    2611.16905

    Read paper →

AI workflows

1. An HVAC and plumbing owner runs his 30-truck company through a team of Grok Bots · Workflow page

  • Replaces: chasing quote follow-ups, building and revising boiler quotes by hand, keeping up with every service call, and dispatching a recall install with “6 texts and a phone call” in the middle of other work.
    • The owner says he gets about 3 times the work done in half the time, mostly away from his desk.
  • What to connect: Grok Bot, one bot per job (Sales, Inbox, HR, Reviews, Ops/CRM, Dash for numbers, Amp for social), plugged into ServiceTitan, email, calendar and text.
    • Quotes are checked against supplier pricing before they go out.
    • Rule: nothing goes out in his name unless he wrote, edited or approved it.
  • Who it’s for: owners of trades and field-service businesses (HVAC, plumbing, installers, contractors) with crews on the road and a busy phone.

2. Inside an automerge setup that went from 809 to 10,926 merged PRs a month · Workflow page

  • Replaces: sitting in a code-review inbox approving pull requests one by one, plus older one-off CI automations.
    • Every PR gets an agentic risk review plus Bugbot; low- and medium-risk ones merge automatically, and only higher-risk ones reach a person.
  • What to connect: Grok Bot (Slack bots with routines) plus Cursor projects/cloud agents, Bugbot and Buildkite.
    • Verification skills: run the app, screenshots or video as proof, fuzzing, property tests, sometimes TLA+.
    • Each repeated agent mistake becomes a check in the codebase.
    • A release bot does a dry run, DMs PR authors, kicks the build and starts a fuzz swarm of 10+ agents.
  • Who it’s for: small engineering teams and solo developers who run coding agents and want more output without reviewing every PR by hand.

3. Weekly train-ticket booking handed to Grok Bot: 30–40 minutes down to 5 · Workflow page

  • Replaces: the Sunday routine of checking next week’s schedule, timetables, prices and railcard discounts, then booking the usual trains.
  • What to connect: Grok Bot with calendar access, past booking history, and the rail operator login plus a payment card stored in the Bot vault.
    • It asked only two questions (login and card) the first time.
  • Who it’s for: commuters and anyone who books the same recurring travel every week.

Quick hits

  • Starship caught by its own tower — On Flight 15, Ship 42 spent about 22.5 hours in space, flew 14 orbits and deployed 26 Starlink V3 satellites.
    • It was caught by Mechazilla’s arms at Starbase around 22:20 Warsaw time (CET) on 14 November. It’s the first ship catch ever.
    • The Super Heavy booster had diverted to the Gulf after a low-fuel alarm and exploded before hitting the water. SpaceX · Space.com
  • “No AI Slop” manifesto — Written at the AIE Code Summit and presented by Dex Horthy, it’s modelled on the Agile Manifesto: “outcomes over output, understanding over speed”.
    • It defines slop as work that cost its maker less effort than it asks the reader to spend.
    • One pledge: say when an agent wrote what you’re sending, and check the parts that matter. It had 391 signatures by 14 November. slopmanifesto.com · @dexhorthy
  • NEJM AI: the physical ceiling on replacing doctors — Stanford, Berkeley and Census Bureau researchers sorted about $45B of 2023 Medicare Part B payments (589,648 physicians, 56 specialties).
    • 24% pays for hands-on work AI can’t do, 21% for work it could, and 55% for visits where the outcome depends on deployment.
    • Radiology is 96% exposed and pathology 99%. Dermatology bills 68% for procedures. Stanford FSI
  • Grok Bot gets cheaper to run, and Grok gets one app — SpaceXAI’s Sam Sokolin says this week’s harness optimizations give about 7% more usage per plan, and up to about 15% more on routine runs.
    • Separately, image and video creation now live inside Grok as part of “one Grok experience”. @SamSokolin · @grok
  • vstack: agents writing formal proofs for inference engines — UCSD’s Hao AI Lab released a 5-stage verification workflow, available as a Claude Code and Cursor plugin.
    • It pins specs by hash so agents can’t weaken invariants to make proofs pass.
    • Applied to prefill-decode disaggregated serving, it found three classes of real RDMA and KV-cache concurrency bugs across SGLang and vLLM, including silent KV corruption that still returns HTTP 200. Hao AI Lab · github.com/hao-ai-lab/vstack

Go deeper

Get the free daily brief

↑