OpenAI files the regulator playbook;
Robinhood opens the broker.

The first major-lab public compliance map. OpenAI's Frontier Governance Framework takes the internal Preparedness Framework — the document that decides whether a model crosses a "high" capability threshold and triggers additional safeguards — and translates it into the artefacts California's Transparency in Frontier AI Act (SB-53) and the EU AI Act's Code of Practice for General-Purpose AI will be asking for starting January. The framework covers cyber offense, CBRN, harmful manipulation and loss of control across model reporting, security risk management, incident response, external expert input and a stated cadence for updates. The strategic read is straightforward: SB-53 takes effect January 1 and the EU GPAI obligations bite the same quarter; whoever sets the template that regulators copy-paste into rulemaking shapes how every frontier lab discloses capability. Anthropic published its own Frontier Compliance Framework in April; OpenAI's answer arrived 41 days after the model that prompted it. Reading the two side by side is now the cheapest pre-rulemaking advocacy briefing in the field.

Published the same Friday as the governance framework, and arguably the more practitioner-relevant of the two. OpenAI's companion paper makes a public case that the eval methodologies inherited from the chatbot era — refusal accounting, public benchmark scores, "LLM as a judge" — break down when the system under test is a long-running agent. The fixes it advocates for: harness design that mirrors real deployment (not the model in isolation), separated private task sets to prevent training-time memorisation, and an explicit norm that third-party evaluators stress-test safeguards before a model crosses a new risk tier. METR and Apollo Research already operate close to this template; the value of the paper is that OpenAI is now publicly endorsing it, which gives buyers something to point at when their procurement team asks why "the model scored 92% on MMLU" doesn't answer "is this agent safe to deploy?" Read with Snorkel's $100M Series D (item 08), the agent-evals category just got both a regulatory tailwind and a capital flywheel inside the same week.

The first US-listed broker to expose its primitives to AI agents through MCP. Robinhood's 27 million customers can now create a separate agentic trading account with a ring-fenced wallet, point an agent at the Robinhood MCP server, and let it analyse the portfolio and place orders against the pre-loaded balance — stock trading only at launch, with options, crypto, futures and prediction markets queued. Alongside it: an agentic Gold Card that issues a separate virtual card number to the agent (deletable any time, 3% cash back, on the same banking MCP), so an autonomous shopper can spend without ever touching the customer's real card. The deal with the agent is enforced at the wallet layer, not the model layer — exactly the boundary Catena Labs's national trust bank charter (covered last week) was designed to formalise. Read alongside the Catena charter filing and Anthropic's split agent-billing credit pool, the agent-finance stack now has three live primitives in production: a wallet, a card, and a settlement lane. The model can spend money now.

Three releases inside 36 hours, finishing the Opus 4.8 launch arc. v2.1.156 was a hotfix for an Opus 4.8 case where thinking blocks were being modified mid-turn, causing API errors — the kind of bug you only find at scale, fixed inside a day. v2.1.157 is the substantive one: plugins inside .claude/skills directories now auto-load without a marketplace, a new claude plugin init <name> scaffolds a plugin in place, slash-command and plugin autocomplete picks up substring matching, the agent field in settings.json is now honoured for dispatched sessions (with --agent <name> to override), and Claude-managed worktrees are left unlocked on finish so git worktree remove actually works. Saturday's v2.1.158 is the regulated- buyer punch line: Auto mode — the classifier-gated tool-call mode that runs allowlisted calls without prompting — is now available on Bedrock, Vertex and Foundry for Opus 4.7 and 4.8 (opt in via CLAUDE_CODE_ENABLE_AUTO_MODE=1). The cloud-provider parity the regulated buyer had been blocked on is now shipped. The classifier reviews each tool call before execution, blocks the dangerous ones, lets the safe ones pass — the operational primitive Cursor 3.6 (item 05) names differently and ships at the same time.

Cursor's response to the same problem Anthropic shipped Auto mode for, and arguably a sharper one. Auto-review is a new run mode that lets Cursor work for longer with fewer approval prompts: Shell, MCP and Fetch tool calls on the allowlist run immediately; calls that can be sandboxed run in a sandbox; everything else goes to a classifier subagent that decides per call whether to allow it, try a different approach, or escalate to the human. The classifier itself is steerable — you can hand it custom instructions about what "safe" means in your codebase. Configured at Settings → Cursor → Agents → Run Mode; landed alongside a Jira integration polish that now posts the resulting pull request URL back into the work item when the agent finishes. Three reads. (1) The cloud-wide Auto mode shipped by Anthropic on Saturday and Cursor's Auto-review shipped on Friday converge on the same primitive: a small classifier in the loop is now what unsafe-by-default agent autonomy looks like in 2026. (2) Cursor is shipping the harness pattern at the IDE; Claude Code is shipping it on the cloud provider's inference. (3) The shared playbook OpenAI argued for the same day (item 02) is going to need to formalise how regulators audit this kind of in-the-loop classifier.

The third leg of the same operational stool. Codex on Windows now supports Computer Use — the agent can see, click and type inside Windows desktop applications while it tests, debugs and refines what it's building, the equivalent of Anthropic's Mac-first Computer Use feature extended to the OS that runs the regulated enterprise. Available in the Codex app today; not available in the EEA, UK or Switzerland at launch (the same SB-53/EU-GPAI dynamic the Frontier Governance Framework is trying to unblock). The second half of the release is more interesting for agent practitioners: ChatGPT on iOS and Android can now start, monitor and steer a Codex session running on a Windows host, with the Windows machine retaining project files, shell, app server and local context. The agent runs on the host you trust; the controller fits in your pocket. With Cursor's classifier and Anthropic's cloud Auto mode shipping the same weekend, the three reference harnesses now all assume the developer is at least sometimes somewhere else.

The Python agent middle layer keeps printing — and noticeably faster than the harness vendors above it. v1.104.0 landed Opus 4.8 support on the stable train along with a Bedrock single-tool tool_choice cache fix and an aligned silent-drop for thinking=False across the hybrid OpenRouter/xAI/Bedrock routes. Twenty-four hours later v2.0.0b4 pulled the same Opus 4.8 support into the harness-first v2 beta line, turning the v1 deprecation warning for prepare-callback None returns into a hard TypeError — the V2 rewrite is now actively refusing the V1 compromises it used to tolerate. Two takeaways. (1) The fastest way to ship Opus 4.8 support across the Python ecosystem in 2026 is to commit it to Pydantic AI and wait — five releases (v1.103, v1.104, v2.0.0b2, b3, b4) hit inside ten days. (2) Anthropic-cadence and Pydantic-AI-cadence have de-facto coupled: when the model ships, Pydantic AI ships the same week.

Addition led the round with Prosperity7 Ventures, existing investors Greylock and Lightspeed, and strategic investors BNY and QBE Ventures — bringing Snorkel's total to $237M. CEO Alex Ratner framed the thesis bluntly: "specialized enterprise agents aren't ready for production in most settings." The round funds the GA of Snorkel Evaluate (programmatic benchmark dataset creation, specialised evaluators, error-mode correction — the explicit alternative to off-the-shelf "LLM-as-a-judge") and Snorkel Expert Data-as-a-Service for advanced-reasoning and agentic-tool-use datasets sold directly to LLM developers. The frontier labs are already customers. The throughline this week is hard to miss: on the same Friday OpenAI published the third-party-evaluations playbook (item 02), Snorkel raised a unicorn round to productise the work that playbook is asking for. The agent-evals category just went from a research debate to a priced market inside one news cycle.

Asana paid $75M for StackAI, the no-code AI workflow platform co-founded by MIT PhDs Tony Rosinol and Bernard Aceituno. The fit is the obvious one — and exactly the one Asana CEO Dan Rogers used to justify the deal. Asana's AI Teammates could plan and track work, but they couldn't execute it end-to-end inside ERP, CRM, ITSM and the rest of the systems where the work actually happens. StackAI's connectors and visual agent-flow builder fill exactly that gap: Teammates now pull context from Asana's Work Graph into a StackAI workflow and push the resulting actions and data back. The strategic read is that "agent-builder" as a standalone category is being consolidated into the workforce-OS layer that owns the user — much as Tonkean went to Coupa earlier in the month (covered then). Expect more: Smartsheet, Monday and ServiceNow all have the same gap, and a $75M cheque looks small against what enterprise buyers will pay to make AI Teammates do the work instead of just describing it.

Capchase's largest round to date — $174M in debt warehouse facilities and $26M in equity led by 01 Advisors, with Caffeinated, Thomvest, Scifi VC, Bling and Invesco. The product the capital underwrites is the part agent practitioners should care about: an Agentic Lending Coordinator that takes the quotes, purchase orders and email chains that used to feed a finance manager's eight-hour packaging exercise, converts them into an executable loan package, and coordinates the multi-party review-through-signing flow between vendor, channel partner and buyer. The headline metric the company is publishing alongside the round: an eight-hour package now takes sixty seconds; 97% of applications are vetted and approved in under thirty seconds. With Catena's national trust bank charter still pending OCC review and Robinhood's MCP wallet now live, this is the first piece of agent-finance infrastructure with a real GTM motion and named enterprise customers (Barracuda, Okta, Verkada, CDW, Insight, Motive). The agent-economy stack just acquired a credit-underwriting primitive.

The biggest European agent move of the year so far. At its "AI Now Summit" in Paris on Thursday, Mistral relaunched Le Chat as Vibe — a single autonomous-work product in four tiers (Free / Pro €14.99 / Team €24.99 per user / Enterprise on request) that combines a chat assistant, a "Work Mode" that hooks into Google Workspace, Outlook, SharePoint, Slack and GitHub to scan inboxes / pull spreadsheets / build reports / push to Notion, and a "Code Mode" that spins up agents in isolated cloud environments to write code, fix bugs and open PRs in parallel. Announced alongside Vibe: Mistral for Industrial Engineering, a stack combining the company's language models with physics simulation from the recently-acquired Emmi AI team, and a new 10MW inference data centre at Les Ulis near Paris. Airbus, BMW and ASML were named as the launch customers for the industrial stack. Three reads. (1) Mistral has explicitly given up "generalist B2B chatbot" as a category and is now positioning as a vertical, sovereign, full-stack European competitor to OpenAI and Anthropic. (2) The Vibe tiering reads like a direct strike at Claude Team and ChatGPT Team on pricing. (3) Airbus/BMW/ASML is the most consequential customer list a non-US lab has produced this year.

Microsoft quietly added Mistral Medium 3.5 to the Copilot Studio model picker the same day the Vibe rebrand landed — worldwide, in early-release environments, with an admin opt-in via the Microsoft 365 admin centre and external providers enabled in the Power Platform admin centre. The model itself is built for the agent shape: "long-horizon tasks, calling multiple tools reliably, producing structured output downstream code can consume," with per-request reasoning effort and explicit EU in-region data residency. Two reads. (1) Copilot Studio has been an OpenAI-and-Anthropic duopoly since the agent feature shipped; Medium 3.5 is the first outside frontier model in the lineup, which lowers the geopolitical risk premium European buyers had been paying. (2) Microsoft listing Mistral on the same Thursday Mistral lists Airbus, BMW and ASML is the closest the European agent stack has come to a coordinated launch — three things pointing at the same European buyer who had been told by procurement to "find a non-US frontier option" inside Q2.

The European push answers the European competitor. Anthropic opened a Milan office on Wednesday — its sixth European location after London, Dublin, Paris, Zurich and Munich — and used the announcement to publish the EMEA run-rate-revenue number that explains the urgency: 9× year-over-year, with large enterprise accounts in the region up 10×. Named Italian launch customers include JAKALA (Claude deployed across 3,000+ users), Satispay and Bending Spoons — both running Claude inside their engineering workflows. Two things make this more than a press-release office opening. (1) The office opening followed Pope Leo XIV's Magnifica Humanitas encyclical on artificial intelligence, at which Anthropic co-founder Chris Olah was invited to speak — the kind of soft-power moment a European challenger like Mistral cannot reproduce. (2) "9× EMEA growth" is the disclosure that lets buyers benchmark Anthropic's European momentum against Mistral's launch-customer list (item 11) and Microsoft's Medium 3.5 inclusion (item 12). The European agent market is no longer a side story; it is competitive front-page news three editions in a row.