The AI Enthusiast.

Claude Code's managed-agent ship.

The Code with Claude 2026 developer event landed three architectural primitives directly into managed agents: multi-agent orchestration, Dreaming (between-turn self-improvement), and Outcomes (rubric-based self-evaluation). The shape of "agent" just got more opinionated.

Claude Opus 4.7 and the quiet generosity

Claude Opus 4.7 sits underneath all of this, generally available since April 16 and now the default for the deeper Claude Code runs. On a 93-task internal coding benchmark, Opus 4.7 lifted resolution 13% over Opus 4.6, including four tasks neither Opus 4.6 nor Sonnet 4.6 could solve. It is also the first Claude model with high-resolution image support (up to 2576px / 3.75MP), keeps the 1M token context, and supports 128k max output.

The under-reported sweetener: Anthropic doubled the Claude Code 5-hour usage limit for Pro, Max, and Enterprise customers. The math only works because Anthropic announced it is partnering with SpaceX's Colossus data center for capacity. Heavier daily limits on top of a stronger frontier model is a meaningful product change disguised as a footnote.

Codex moves into your pocket.

On May 14, OpenAI brought Codex preview into the ChatGPT mobile app for iOS and Android. The agent runs on your Mac; the phone becomes the cockpit — start or continue threads, approve actions, review diffs, screenshots, and test results from anywhere. This is the most "actual mobile coding" any major lab has shipped.

The important OpenAI move this week is not a benchmark bump. It is a workflow change. On May 14, 2026, OpenAI made Codex mobile-native enough that a phone can keep an active coding thread moving while the actual work continues on a connected Mac or remote machine. Around that surface, OpenAI is hardening the operational layer: Remote SSH is now generally available, Hooks are generally available, and programmatic access tokens open the door to governed automation in ChatGPT workspaces.

MCP crosses the enterprise chasm.

Model Context Protocol won the connectivity argument. The conversation has shifted entirely to operating MCP servers reliably at enterprise scale: governance, observability, outcome metrics, and runtime control.

• Central MCP server management dashboards are emerging as a discrete product category — per-tool RBAC, audit trails, and outcome-based evaluation built in.
• Outcome-based agent eval is replacing vibes-based "did the output look useful?" with task success, tool correctness, latency, retries, policy compliance, escalation quality, and cost per successful outcome. Anthropic's Outcomes feature (Page 04) is the same idea formalized.
• Adobe CX Enterprise launched as a marquee example: agentic AI system bringing together AI agents, skills, and MCP endpoints across the customer lifecycle — full marketing workflows wrapped around governed MCP.
• Authentication is the bottleneck. Multiple analyst pieces this month make the same point: as MCP rolls out, the unlock is "powerful and user-friendly interfaces to establish and authenticate connections," not more models or more tools.

The takeaway for any team running production agents: stop measuring on vibes. Define your outcomes rubric this week, instrument it, and treat the MCP runtime as a first-class operational concern, not a glue layer.

EU AI Act simplified — 79 days out.

On May 7, 2026, the European Parliament and Council reached a political agreement to simplify the AI Act. The bulk of the Act still becomes applicable August 2, 2026. That is 79 days from this issue.

The "Brussels effect" is also still operating: the AI Act is functioning as a global reference standard the same way GDPR did, with US products voluntarily adopting EU-style compliance even where they have no legal obligation. If your product touches the EU and you have not done a high-risk classification review under the new narrower scope, you are likely now compliant by default — but you still need to document it.

Anthropic's $950B price tag.

Anthropic is reportedly in talks to raise $30–50 billion at a $950 billion valuation — within shouting distance of the $1T mark held only by Nvidia, Microsoft, Apple, Google, and Amazon. And the enterprise contracts justifying it kept landing this week.

• Gates Foundation announced a $200M partnership applying Claude to global health and development.
• PwC is deploying Claude across its consulting practice — to "build technology, execute deals, and reinvent enterprise functions for clients."
• Claude Platform on AWS reached general availability — full Claude Platform features with AWS authentication, billing, and commitment retirement.
• Claude for Excel, PowerPoint, and Word are now generally available; Claude for Outlook is in public beta for all paid plans — direct shot at Microsoft Copilot's home turf.
• Claude for Small Business launched with connectors for QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365 — more on Page 10.
• Anthropic legal tools added 12 new practice-area plugins covering corporate, regulatory, and employment law.

The valuation looks aggressive in isolation. Stacked against $30B+ run-rate revenue, 1,000+ customers spending $1M+ annually, the SpaceX/Colossus capacity deal, and the new SMB on-ramp, it is a coherent (if exuberant) bet. The risk is no longer "can Anthropic ship?" — it is "can compute supply, hiring, and enterprise services keep up with the demand they have already booked?"

Claude gets a small-business job.

On May 13, 2026, Anthropic launched Claude for Small Business: a package built around the tools owners already use, including QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365.

The product is not pitched as a blank chatbot. Anthropic is packaging repeatable work: payroll planning, invoice follow-up, month-end accounting prep, contract review, sales campaigns, marketing assets, customer responses, and HR admin. The more important design choice is that owners still approve key actions before anything is sent, posted, paid, or filed.

The newsletter takeaway: this is Anthropic's clearest Main Street move yet. Small businesses are economically huge but adoption-constrained by time, confidence, and lack of IT staff. Anthropic is betting that AI adoption accelerates when the product starts with the admin backlog owners already hate.

OpenAI's TanStack warning: update your Mac apps.

On May 13, 2026, OpenAI said two employee devices were impacted by the TanStack npm supply-chain attack. OpenAI found no evidence that user data, production systems, intellectual property, deployed software, or existing installs were compromised, but it is rotating code-signing certificates and telling macOS users to update by June 12, 2026.

The item belongs in this issue because it happened after last Friday's May 8 briefing and inside the May 9-15 coverage window. It also argues for a standing newsletter slot: every week needs a security section that checks model/provider incidents, package supply-chain attacks, app update advisories, AI abuse reports, and policy-relevant vulnerability news.

What to watch before next Friday.

This issue closed with the open-source agent stack visibly specializing, Google setting the table for May 19, and OpenAI turning Codex into a cross-device surface. The next turn is about what becomes real over the next seven days.

One practical note: coverage window for this issue is May 9, 2026 through May 15, 2026. The most date-sensitive correction is Google I/O itself: as of publication, the event has not happened yet.

That is the briefing. See you next Friday.

Built from this week's roundup, then corrected against primary sources and exact dates. The main fix in this issue: Google I/O coverage is framed as a pre-keynote packet because the conference starts on May 19, 2026, four days after publication.