Agents moved from demos toward infrastructure this week. OpenAI shipped a dedicated endpoint for fast agent decisions, Anthropic gave the Agent Skills standard a neutral home at the Linux Foundation, and Google Cloud announced the microVM service it says already runs Gemini’s agents.
Trust and control were the other thread. Microsoft argued that companies should treat their own agents as potential insiders, Visa, BILL and Billtrust put an agent into a live payments pilot that only executes buyer-approved invoices, and Anthropic and Theorem used Claude to write machine-checked proofs for open-source code.
On the Grok and Cursor side, SpaceXSI launched an official partner network for companies building with Grok and described two Grok Bot teammates its support team uses during incidents. There was no Oct 22 daily edition; the Oct 23 edition covered 48 hours.
TL;DR
8 takeawaysOpenAI made the Decisions API generally available, built with GPT-6 Luna and priced at $0.10 per million input tokens with no output-token charges
SpaceXSI launched the SpaceX Super Intelligence Partner Network for Grok, saying more than 130 companies took part in early access
Anthropic donated Agent Skills to the Agentic AI Foundation, saying more than 70 agent platforms have integrated it
Anthropic and Theorem say Claude proved no-crash properties for 239 entry points across 123 open-source libraries
Visa, BILL and Billtrust launched a live pilot in which an AI agent pays approved supplier invoices
Microsoft’s security chief urged companies to treat their own AI agents as “agentic insiders”
SpaceXSI described Grok Bot teammates that it says cleared 204 incident tickets in about 10 minutes
Notion rebuilt its agents to run tasks for up to 24 hours and added a cheaper Auto model router
Top Stories of the Week
Story 01 / 08 · Thu, Oct 22
OpenAI makes the Decisions API generally available
- What happened
- OpenAI made the Decisions API generally available. Built with GPT-6 Luna, it answers questions with bounded answers, returning a probability, a choice from supplied options, or a rubric score from text, images or native JSON. OpenAI says it is up to 10x faster than using GPT-6 Luna through the Responses API. Since the public beta, OpenAI added a native GPT-Live integration, prompt caching, US and EU inference residency, and up to 400 questions per request. Pricing is $0.10 per million input tokens, $0.01 cached, with no output-token charges.
- Why it matters
- Builders: Routing, classification and next-action choices can move to a cheaper, faster endpoint. OpenAI quotes Datadog reporting 40x faster p95 than available frontier models and Sierra running under 280 ms p95; test on your own decisions.
- Enterprises: US and EU regional processing and Zero Data Retention for eligible customers, per OpenAI, make it easier to use in regulated workflows.
- Startups: Input-only pricing makes high-volume decision calls predictable. Ramp told OpenAI it was 4x faster and 30% cheaper than GPT-6 Luna for attributing AI costs.
- In the daily digest
- Oct 23 edition
Story 02 / 08 · Thu, Oct 22
SpaceXSI launches a partner network for Grok
- What happened
- SpaceXSI introduced the SpaceX Super Intelligence Partner Network, its official program for companies building products, solutions and services with Grok. Transformation Partners help customers deploy the Grok platform across the enterprise, and Marketplace Partners build plugins and integrations. SpaceXSI says more than 130 companies took part in early access and that engineers at more than 30 organizations have been trained through its Forward Deployed Engineer program. Quoted partners include Bain & Company, Databricks, NVIDIA and Snowflake.
- Why it matters
- Builders: Tool vendors can apply as Marketplace Partners to ship integrations for the Grok product suite.
- Enterprises: Companies adopting Grok now have an official list of accredited delivery partners.
- Startups: A marketplace route puts smaller vendors alongside large consultancies in front of Grok customers.
- In the daily digest
- Oct 23 edition
Story 03 / 08 · Thu, Oct 22
Anthropic donates Agent Skills to the Agentic AI Foundation
- What happened
- Anthropic is donating Agent Skills, the open standard in which a folder with a SKILL.md file teaches an agent a task, to the Agentic AI Foundation under the Linux Foundation. Anthropic co-founded the foundation with Block and OpenAI and previously donated the Model Context Protocol to it. The donation includes the name, logo, GitHub organization and agentskills.io. Anthropic says more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, have integrated the standard, with 1 million skills published and nearly 280 million installed.
- Why it matters
- Builders: Skills written once can be reused across agent platforms under vendor-neutral governance.
- Enterprises: Neutral stewardship reduces lock-in risk for internal skill libraries.
- Startups: Anthropic says the roadmap includes versioned releases, a proposal process and security guidance for skill authors.
- In the daily digest
- Oct 23 edition
Story 04 / 08 · Thu, Oct 22
Claude writes machine-checked proofs for 123 open-source libraries
- What happened
- Anthropic and Theorem report that Claude proved 239 OSS-Fuzz entry points across 123 open-source libraries always exit normally without memory-safety errors or failed assertions, for every input under 64 MB. The proofs operate on compiled RISC-V binaries and total about 180 million lines of Rocq, according to Anthropic, and Claude found and patched bugs along the way. The guarantees assume a single-threaded static binary with a fixed memory budget on a RISC-V CPU running Linux.
- Why it matters
- Builders: Theorem is offering to run the pipeline for other open-source projects and publishing a Claude skill to get started.
- Enterprises: Proof-backed guarantees could complement testing and fuzzing for security-critical code, within the stated assumptions.
- Startups: Anthropic calls it, to its knowledge, the first large-scale AI effort at assembly-level verification of production software, a possible new category of security tooling.
- In the daily digest
- Oct 24 edition
Story 05 / 08 · Tue, Oct 20
Visa, BILL and Billtrust put an AI agent into live B2B payments
- What happened
- Visa, BILL and Billtrust launched a live pilot in which an authorized BILL Payment Agent, registered through Visa’s Trusted Agent Protocol and Agentic Directory, pays approved supplier invoices with tokenized card credentials. Billtrust uses the same protocol to recognize and authenticate the agent on the supplier side. The companies stress that the agent does not make payment decisions; it executes payments the buyer has already approved.
- Why it matters
- Builders: Agent identity protocols, not CAPTCHA workarounds, are becoming the way agents get through payment portals.
- Enterprises: Accounts payable teams no longer need to share supplier-portal logins with automation, according to the companies.
- Startups: Fintech builders should watch Visa’s Trusted Agent Protocol as a registration path for their own agents.
- Source
- Visa(opens in a new tab)
- In the daily digest
- Oct 21 edition
Story 06 / 08 · Mon, Oct 19
Microsoft: treat your own AI agents as “agentic insiders”
- What happened
- On the Official Microsoft Blog, Microsoft Security EVP Hayete Gallot argued that agents deployed inside companies are often overprivileged and may not behave as intended even without being compromised, and called them “agentic insiders.” Microsoft says Project Perception can discover agents, assign each an identity and permissions, compute risk dynamically and use Guardian Agents to supervise actions in real time.
- Why it matters
- Builders: Design agents with least privilege and per-agent identity from the start.
- Enterprises: Gallot says protection must span identity, permissions, data, endpoints, network, cloud and runtime, with continuous monitoring.
- Startups: Agent discovery and supervision is shaping up as a security category large vendors are targeting.
- In the daily digest
- Oct 19 edition
Story 07 / 08 · Tue, Oct 20
SpaceXSI shows Grok Bot teammates handling support incidents
- What happened
- SpaceXSI described two Grok Bot teammates its support team uses during product incidents. IncidentWatch reviews the previous 15 minutes of tickets every five minutes and alerts the on-call support engineer to spikes. IncidentSweep finds tickets tied to a resolved incident, shows an engineer the proposed batch and message for approval, then sends in small batches. The authors say it cleared 204 tickets in about 10 minutes on September 18.
- Why it matters
- Builders: The pattern keeps a human approval step and stops and reports on any failure.
- Enterprises: Support teams can apply the same approve-then-send design to incident follow-ups.
- Startups: A concrete, measurable workflow is a useful template for selling agent automation.
- In the daily digest
- Oct 20 edition
Story 08 / 08 · Mon, Oct 19
Notion rebuilds its agents and adds a cheaper model router
- What happened
- Notion says it rebuilt its agent harness: Notion Agents get their own computer and browser, smarter context management and a “Review for me” mode, and can work on tasks for up to 24 hours. Notion says its upgraded Auto model router delivered results similar to Opus 5.5 and Astra at about 90% lower average cost across 20+ common tasks. Notion also overhauled its MCP server so outside agents can use Notion as shared memory.
- Why it matters
- Builders: Notion via MCP becomes a shared memory layer for agents built elsewhere.
- Enterprises: All AI usage moves to a new credit system, with an AI Command Center for spending limits listed as coming soon.
- Startups: Model routing for cost is now a feature of mainstream productivity tools.
- In the daily digest
- Oct 20 edition
Also Notable
Block unveils Dessa
Block says its full-duplex speech model was trained only on human speech and runs locally on a MacBook M4 Max or better; no public access yet, TechCrunch reports.
TechCrunch(opens in a new tab)Google Cloud Agent Compute Engine
Google says the hardware-isolated microVM service already powers Gemini’s work agents; preview is planned in the coming months.
Google Cloud(opens in a new tab)Kimi K3.1 Pro
Moonshot AI released K3.1 Pro with a 1M-token context and promises full weights by Nov 4; benchmark claims are Moonshot’s own.
Moonshot AI(opens in a new tab)NVIDIA Nemotron 3.5 Embed, Rerank and Parse
Open retrieval models for agents; NVIDIA says adding Rerank cut agent token cost per query by about 14% with Embed 8B.
NVIDIA / Hugging Face(opens in a new tab)Claude Duet research preview
Anthropic’s proactive agent app, by waitlist starting with US Pro and Max subscribers.
Anthropic(opens in a new tab)Google Public Sector previews Gemini for government services
In preview for US federal, state and local governments, per Google.
PR Newswire / Google Public Sector(opens in a new tab)OpenAI models coming to Google Cloud
OpenAI says its models and Codex workflows will be available in Model Garden, with general availability planned in the coming weeks.
OpenAI(opens in a new tab)Tenstorrent raises $1.5B Series E
At a $6.2 billion post-money valuation, led by Trace Capital and Innovation X, per Tenstorrent.
Tenstorrent(opens in a new tab)MiniMax M3.1 Flash
MiniMax says it is 70% cheaper than its predecessor for agentic tasks, with weights to come later.
MiniMax(opens in a new tab)Mistral Voxtral 3
Speech-to-text and text-to-speech models plus a LiveKit-based voice agent SDK.
Mistral(opens in a new tab)
Themes of the Week
Agents get infrastructure
Decision endpoints, microVM sandboxes and a neutral skills standard all treat agents as a platform layer rather than a feature.
Humans stay in the approval loop
Visa’s pilot, SpaceXSI’s IncidentSweep and Microsoft’s agentic-insider framing all keep a person approving what agents do.