Machine-Checked Agents Move From Demos to Real Work

Saturday, 14 November 2026 | Global AI News Daily Ledger

Google Research put multi-agent systems to work on open problems and shipped the receipts - a full proof of the Courtade-Kumar conjecture, a 1.8-million-line Lean formalization of the Weil conjectures, and bit-exact Rust rewrites - while opening /teamwork, /boost and Deep Review to external researchers. OpenAI gave agents a way to pay with ChatGPT Wallet, built with the card networks, and folded its stand-alone robotics team into general-model research. The NVIDIA fallout widened on three fronts: a WSJ report alleging circular GPU financing, a lawmaker saying NVIDIA privately confirmed chip tracking is feasible, and a forensic claim that China's top open models carry Blackwell training fingerprints. llama.cpp reached v1.0.0, LangChain raised $300 million at a $3 billion valuation, GitHub Next released its Ace shared workspace, and Nex-AGI shipped a 321B open agentic model.

Coverage window Since 13 November, 18:45 IST | Compiled 14 November, 18:44 IST

1. Google's /teamwork agents prove open math conjectures and rewrite kernels, now open to outsiders

Google Research on 13 November published a sweep of results from Antigravity's /teamwork multi-agent workflows paired with Gemini 4 models, and opened the tools - /teamwork, /boost and Deep Review - to external researchers and developers, with 10,000 academic researchers offered six months of free top-tier access. Highlights: researchers from Google, CUHK, USC and CMU proved information theory's 2013 Courtade-Kumar conjecture in full generality, formalized in Lean 4; a /teamwork-LEAP setup produced an end-to-end machine-checked formalization of the Weil conjectures spanning about 1.8 million lines of Lean across 28,800 subagent invocations, the first formalization of a recent Hilbert's-tenth result over general rings of integers, and reproductions of a finite-time-blowup proof for forced 3D Navier-Stokes. A two-week Argon campaign rewrote the remaining C++ of the Zircon microkernel into about 800,000 lines of safe Rust passing all kernel unit tests, and /teamwork's Rust port of the libgav1 AV1 decoder reached bit-exact parity and then ran 20-30% faster than the original C++ and hand-optimized assembly. Ahead of STOC 2027, authors of 578 submissions opted into Deep Review; in an ongoing survey, over 30% said it caught substantive proof bugs or significant typos.

Analysis. This is the strongest single-day evidence yet that multi-agent harnesses produce verified research and systems output, not demos: every headline claim is anchored to a Lean kernel check, an unmodified test suite, or a hardware measurement. The results are Google's own collaborations, so independent review of the papers matters - but machine-checked proofs are auditable by construction.

https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/
https://antigravity.google/docs/teamwork/

2. OpenAI launches ChatGPT Wallet so agents can pay, with the networks on board

OpenAI on 13 November introduced ChatGPT Wallet, built with American Express, Mastercard and Visa, initially for US ChatGPT Pro subscribers aged 18+ through its dots agents. Users add a card once and approve each purchase with a security check such as Face ID; the payment network then issues a token with a substitute card number, so neither the agent nor the merchant sees the real card details, and existing card rewards and eligible dispute protections carry over. OpenAI says delegated purchases with pre-approved parameters like spending limits and categories are planned. Alongside the wallet, it published a Commerce Partner Playbook telling businesses to recognize signed agent requests via Web Bot Auth, and named fraud and payments partners including Forter, Riskified, Sardine, Signifyd, Adyen, Checkout.com and Stripe.

Analysis. Two days after United moved to block unauthorized agents, OpenAI shipped the authorization layer that answers it: network-issued tokens, per-purchase biometric approval and a paper trail businesses already understand. Agentic commerce is consolidating around the card networks' rails rather than bypassing them, which puts pressure on every 'block the bots' stance to define what an authorized agent is.

https://openai.com/index/introducing-chatgpt-wallet
https://www.theverge.com/ai-artificial-intelligence/1033425/openai-chatgpt-wallet
https://help.openai.com/en/articles/20001654-making-purchases-with-chatgpt-wallet

3. OpenAI disbands its stand-alone robotics team, betting general models can run robots

OpenAI has disbanded its stand-alone robotics team and reassigned its researchers - robotics research to the main research org under Chief Research Officer Mark Chen, and applied robotics and hardware to the Applied organization under President Greg Brockman - the company confirmed to Fortune on 13 November. OpenAI says the move is meant to accelerate robotics work, not retreat from it: its general frontier models, including GPT-6 Astra, have shown they can be applied directly to robot control, making a separate specialized-model team unnecessary. An NVIDIA study cited by Fortune found an Astra-powered agent achieved 74.8% on a simulated manipulation benchmark versus 51.0% for the best specialized models. OpenAI still plans a first public demonstration of its own humanoid robot in 2027. CMU's Guanya Shi read the move as OpenAI treating robotics as 'no longer a special problem' but a downstream application of general intelligence, while noting smaller on-robot models still have a role when devices can't depend on the cloud.

Analysis. If the biggest model lab folds robotics into general-model research, the embodied-AI thesis shifts from specialized brains to general models plus site-specific data - which is exactly where robotics startups like Walden and Generalist say their deployment advantage lives. This is a company reorganization around an early benchmark result, so the 2027 demo is the real test.

https://fortune.com/2026/11/13/openai-disbands-robotics-team-as-its-general-models-start-running-robots/
https://the-decoder.com/openai-reassigns-robotics-researchers-to-other-teams-as-general-purpose-models-might-make-specialized-robotics-ai-unnecessary/

4. WSJ report alleges NVIDIA fueled GPU scarcity by funding startups that buy its chips

A Wall Street Journal report, surfacing after Jensen Huang's resignation and circulated via Wccftech and TechPowerUp, alleges NVIDIA inflated the perception of GPU demand by investing in AI startups and small cloud providers that then bought NVIDIA GPUs - a circular-financing pattern reportedly driven partly by an oversupply of semiconductors coming from TSMC. The report cites an August meeting in which CFO Colette Kress - now interim CEO - pushed to scale back circular financing arrangements, including the $250 billion OpenAI deal. Separately, TechCrunch reports NVIDIA participated in 105 funding rounds this year, often as the largest investor, backing CoreWeave, Crusoe, Mistral AI, Lambda, Black Forest Labs and Reflection AI, and that VCs and founders now expect a pullback under Kress that could tighten early-stage AI funding.

Analysis. Circular vendor financing has been the AI boom's open question, and this is the first reported evidence it was a deliberate demand-management strategy rather than exuberance - but note the sourcing: the account reaches secondary outlets via Wccftech, and the allegations remain unadjudicated reporting. Kress's documented skepticism of these arrangements is now an asset as she runs the company.

https://www.techpowerup.com/354814/nvidia-allegedly-engineered-scarcity-by-funding-ai-startups-to-buy-its-gpus
https://www.explainx.ai/blog/nvidia-engineered-gpu-scarcity-startup-funding-wsj-claimed-vs-verified-2026
https://techcrunch.com/2026/11/13/jensen-huangs-departure-blasts-big-hole-in-ai-startup-funding-vcs-and-founders-say/

5. NVIDIA privately told a lawmaker chip tracking is feasible, despite industry claims

NVIDIA representatives told Representative Bill Foster, cosponsor of the Chip Security Act, that the company could technically implement the bill's anti-smuggling geotracking requirements, Foster told MIT Technology Review in an interview published 12 November - even as the Semiconductor Industry Association, which counts NVIDIA as a member, publicly called such mandates 'new, untested, and potentially infeasible' in March. The act, one of three chip-security bills under consideration for the Senate's National Defense Authorization Act, would require location verification for advanced AI chips through inspections, attestations, or ping-based mechanisms. NVIDIA did not respond to requests for confirmation. Foster says the fallout around Huang's resignation has fueled congressional support, and he plans stricter follow-up legislation next year requiring verified licenses before a GPU can operate.

Analysis. A private 'yes we can' from the company whose trade group publicly said 'potentially infeasible' removes the main technical argument against chip tracking; the debate is now purely about cost and willingness. Combined with the fingerprint research below, export enforcement is shifting from paper rules to technical mechanisms.

https://www.technologyreview.com/2026/11/12/1148483/nvidia-chip-tracking/
https://thedefiant.io/frontier/ai/foster-says-nvidia-privately-deemed-chip-tracking-feasible

6. Forensic analysis claims China's top open models carry Blackwell training fingerprints

Will DePue, a former OpenAI researcher running Unemployment Labs, published an analysis claiming six Chinese open-weight models - including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3 - were initialized on export-restricted NVIDIA Blackwell systems, with four matching GB200/GB300 Grace Blackwell racks. The method exploits the fact that GPU-specific pseudorandom weight initialization (Philox/cuRAND layouts) survives into published weights; DePue says he searched over six trillion chip-layout and seed combinations and published initial seed values, including DeepSeek-V3 on Hopper and DeepSeek-V4-Flash on Blackwell, so others can reproduce the work. He also notes the limits: Blackwell fingerprints do not prove where the GPUs sat, and offshore data centers - 275,000 NVIDIA accelerators entered Malaysia in 2024-25, per Epoch AI - may be legal under current rules. The White House earlier alleged Moonshot used banned chips, and Senators Banks and Warren cited the Huang affair this week in pushing the AI Overwatch Act; the Trump administration says the espionage allegation is unconfirmed and Beijing denies it.

Analysis. If reproducible, this is a new enforcement instrument: open weights become evidence of training hardware. The claim is a forensic inference from one researcher - published seeds invite replication, and the distinction between 'trained on Blackwell' and 'violated export law' is doing real work. China's frontier capability clearly is not waiting for permission.

https://techcrunch.com/2026/11/13/export-banned-nvidia-chips-are-behind-chinas-most-powerful-ai-models-uh-oh-jensen/
https://www.theregister.com/ai-and-ml/2026/11/13/chinese-ai-labs-are-training-models-on-nvidias-blackwell-despite-export-controls-says-ex-openai-engineer/5309891

7. llama.cpp ships v1.0.0 and commits to semantic versioning

The ggml-org team released llama.cpp v1.0.0 and ggml v1.0.0 on 13 November, the project's first stable major release after three and a half years. From here, breaking changes to the public C API trigger major-version bumps, with minor versions for backward-compatible features, backends and model support. The release formalizes what underpins much of local AI: llama.cpp supports more than 150 model architectures and mainstream speculative decoding methods across backends for virtually any hardware, with distribution through pre-release artifacts for Linux, macOS, Windows, Android and iOS plus Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and XCFramework channels. The team credits HF Buckets and HF Jobs for CI capacity and lists hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.

Analysis. Version 1.0 here is not a feature release, it is a stability contract: downstream apps can finally pin to an API that will not break silently. That matters more each month as local and private inference becomes the fallback for people who do not want their prompts on someone else's server.

https://huggingface.co/blog/ggml-org/llama-cpp-v1
https://github.com/ggml-org/llama.cpp/releases/tag/v1.0.0

8. LangChain raises $300M Series C at a $3B valuation around LangSmith

LangChain on 13 November announced a $300 million Series C at a $3 billion valuation, led by CapitalG with new investors GIC and DFJ Growth plus IVP, Sapphire, Salesforce Ventures, Insight, QuantumLight and MongoDB Ventures, and Sequoia, Amplify and Benchmark doubling down. The pitch centers on LangSmith, its framework-agnostic platform to run, improve and govern agents: more than 7,000 organizations use it, including NVIDIA, Toyota, Salesforce, Workday, Harvey and Rippling, with Clay monitoring hundreds of millions of agent runs per month and Workday cutting new-agent deployment time from five months to under a day. LangChain framed the raise around agent runtime, improvement and governance pillars, and teased new LangSmith capabilities for a 20 November event.

Analysis. The agent-tooling layer is consolidating into real enterprise infrastructure budgets, and the customer list shows governance and observability - not frameworks - are what buyers pay for. Note the timing contrast: this landed the same week NVIDIA's pullback started squeezing the early-stage end of the market.

https://www.langchain.com/blog/series-c
https://techcrunch.com/2026/11/13/ai-agent-startup-langchain-hits-3b-valuation/
https://capitalg.com/insights/doubling-down-on-langchain/

9. GitHub Next releases Ace, a local shared workspace for humans and coding agents

GitHub Next has released Ace, an open-source collaborative coding environment where people and AI agents work together in channels running on their own machines. Teams connect over Tailscale, supply their own Anthropic or OpenAI API keys, and need no separate accounts or invitation system. The trust model is explicit and worth reading before adoption: agents run with the host user's privileges, can execute arbitrary shell commands and access files beyond the project directory, and channels and lanes are not security sandboxes - the Tailscale tailnet is the only membership and access boundary.

Analysis. Ace is the local-first answer to cloud agent platforms: your keys, your machines, your tailnet. The same design makes the blast radius obvious - an agent with your privileges and no sandbox means prompt injection inside a shared channel is a team-level incident, so treat channel membership like production access.

https://github.com/githubnext/ace
https://aibriefing.imbch.dev/en/article/11923505
https://dev.to/piekwerk/github-nexts-ace-runs-agents-with-your-privileges-7-facts-from-its-trust-model-and-6-checks-before-you-share-a-channel-53hh

10. Nex-AGI ships Nex-N2.6, a 321B open agentic model built on GLM-5.3-Flash

Nex-AGI released Nex-N2.6 on 14 November, a roughly 321-billion-parameter multimodal agentic model built on GLM-5.3-Flash under Apache 2.0, with FP8-quantized weights on Hugging Face. The model card reports 83.0 on OSWorld versus 78.0 for the base GLM-5.3-Flash and 84.1 on Terminal-Bench 2.1, plus an internal checkpoint-replay evaluation in which reasoning repetition fell from 16.0% to 1.6% versus Nex-N2.5-Pro. Nex-AGI also claims 100% tool-list adherence among responses making tool calls - a figure from its own benchmark set, not an external audit. AI Weekly notes the HF page recorded 410 downloads in the past month.

Analysis. A 321B Apache-2.0 agentic model on a GLM base gives agent builders another open option above the laptop tier, and the OSWorld gain over the base model is the number to watch. The 100% tool-adherence claim is self-measured and should be treated as marketing until replicated.

https://aiweekly.co/alerts/nex-agi-releases-nex-n26-321b-agentic-model-on-glm-53-flash

Scope and sourcing notes

Ten distinct underlying developments. Google's /teamwork post covers new research results plus external availability; the feature's August announcement is not repeated. ChatGPT Wallet is rolling out to US ChatGPT Pro users aged 18+ through dots, with per-purchase approval; broader delegated purchasing is a stated plan. The WSJ circular-financing account reaches secondary outlets via Wccftech and remains unadjudicated reporting; NVIDIA has not publicly responded to it in sources checked. The Blackwell fingerprint analysis is a forensic claim from one researcher: seed values are published for reproduction, GPU location is not established, and no export violation is proven. Nex-AGI's benchmarks and 100% tool-adherence figure are self-reported. GitHub Next's Ace runs agents with host-user privileges and no sandbox; its trust documentation is linked. OpenAI's robotics reorganization is company-confirmed, and the 2027 humanoid demo remains a plan. Weighed and dropped: the TechCrunch NVIDIA funding-gap reaction piece (covered within story 4), Gemini glasses preorders, Hydrogen's GitHub Copilot GA, Claude Code's Code Mode, the Scott Wiener profile, OpenAI's Shield and Amplitude case studies, and Unsloth's quantization update. No public X post is used as a Firsthand signal today.

Bring agentic AI to your classroom

TechSambad's AI Learning Pilot puts agentic AI in front of your students for one module, not a semester: a live 90-minute taught session inside your timetable, an AI-built pre-class reading brief mapped to your module, a question bank with grading rubrics, a student takeaway pack, and a written outcome report. Run it once and judge it on what your students can do afterwards - Rs 2,000 + GST for the college, no obligation.

https://www.techsambad.com/p/ai-learning-pilot.html?m=1