Claude writes 80% of Claude —
Anthropic asks for a global brake pedal.

The cleanest first-party admission of the year, and the one Anthropic chose to make on the same week the Ramp index has it overtaking OpenAI in US business adoption (item 02). On June 4, Anthropic published "When AI Builds Itself", co-authored by co-founder Jack Clark and the head of its internal policy institute, Marina Favaro. The disclosures are specific and they all point the same direction. More than 80% of the code merged into Anthropic's own systems is now authored by Claude, up from low single digits before Claude Code launched in early 2025; engineers ship roughly 8× the code per day they did in 2024; and the time horizon over which an AI agent can run autonomously on a task is doubling roughly every four months. The framing is careful: Anthropic is explicit that recursive self-improvement has not arrived — "we are not there yet, and recursive self-improvement is not inevitable" — but writes that "it could come sooner than most institutions are prepared for," and asks for the world to keep ready the option to slow or temporarily pause frontier AI development. Two reads. (1) The disclosure is also a sales document: every CIO comparing Claude Code to competitors now has a first-party datapoint that the lab uses its own product on itself, in production, at a level no rival has matched on the record. (2) The ask — a coordinated international brake pedal — is the second time in three months Anthropic has openly campaigned for regulation that would bind its own roadmap; read with the Project Glasswing disclosures and the Mythos deferment, the policy posture is the brand, and the timing puts it in front of the developer audience in Tokyo (item 03) before any of it has to clear the EU AI Act August 2 cutover.

The data point under the policy memo, and the line every enterprise AI vendor is now graphed against. The Ramp AI Index's May 2026 release, drawn from invoicing and spend data across roughly 50,000+ US companies, reports that Anthropic is now paid for by 34.4% of US businesses, against 32.3% for OpenAI's ChatGPT — the first time Anthropic has led the headline metric since the series began. Month-on-month, Anthropic's share rose +3.8 pts in April; year-on-year, Anthropic has roughly quadrupled its enterprise share while OpenAI's grew by 0.3 pts. Ramp attributes most of the delta to Claude Code, which the index calls the fastest-growing product in Anthropic's history; enterprises cite reliability, long-context handling and instruction-faithful behaviour as the production-workload differentiators. Two reads. (1) The number is a leading indicator, not a final score: Ramp data only catches Ramp's customer base, and the OpenAI side of the same ledger now includes Microsoft Azure resold tokens not attributed to "ChatGPT," so the gap is directionally real but not as wide as the headline. (2) Pair with the brake-pedal memo (item 01) and the Code w/ Claude Tokyo stop (item 03): Anthropic is making the same pitch — capability, safety, distribution — in three different channels in the same week, and the enterprise data is finally moving with the rhetoric.

The same week Anthropic ships the institute paper, it also ships the keynote. Code w/ Claude Tokyo runs June 5–6 at Anthropic's Roppongi venue, with Day 1 keynotes and breakout sessions streamed live — Anthropic's third stop in the 2026 developer-conference tour after San Francisco and London. The framing the same speaker bench has been using on stage since SF (Cherny: "the default isn't 'I'm going to prompt Claude' — the default is now 'I'm going to have Claude prompt itself'") now reads as a soft preview of item 01: Claude prompting Claude is exactly what Anthropic is disclosing on its own codebase. Two reads. (1) The tour function isn't the keynote, it's the back-room — every previous stop has been the moment enterprise pilots in that region clear into production; Tokyo is the JP/APAC version of the same conversion funnel, and Anthropic's enterprise momentum needs an APAC leg if the Ramp number (item 02) is going to be a global story and not a North American one. (2) Pair with the lead memo: the speakers on stage in Tokyo are the same authors of the policy ask in Washington, and putting both in front of a developer audience inside 48 hours is the cleanest way to make "capability up, safety up, distribution up" a single consistent message.

The cleanest sign of where OpenAI thinks the agent-platform value lives. On June 3, OpenAI confirmed that Agent Builder and the Evals product — the no-code agent canvas and the eval surface shipped at AgentKit's October launch — are being wound down, and will no longer be available on the OpenAI platform from November 30, 2026. The migration path the changelog spells out: code-shaped workflows move to the Agents SDK; chat-shaped, natural-language workflows move to Workspace Agents in ChatGPT; ChatKit remains as the embeddable front end. Two reads. (1) AgentKit was launched as three products bolted together — the SDK, the visual Builder, and Evals — and inside nine months, two of the three are being removed. That's not a clean signal about OpenAI's commitment to agents; it's a clean signal that visual no-code agent canvases have lost the developer mindshare race to typed SDKs, and that evals as a hosted product have lost to evals as a library inside the SDK. (2) Pair with the Ramp number (item 02): the products OpenAI is now discontinuing are exactly the surfaces an enterprise buyer was using to scope an Anthropic alternative — deprecating them mid-procurement-cycle is the kind of policy move that shows up as another negative tick on next month's Ramp index.

The data-warehouse vendors are turning into agent vendors in real time. At Snowflake Summit 26 (June 2, San Francisco, record 20,000+ attendees), Snowflake GA'd Snowflake CoWork — the rebrand of Snowflake Intelligence — as the personal work agent for knowledge workers. Three things matter. (1) Cortex Sense: a new grounding layer that unifies data, business definitions, and operational knowledge into a shared context layer for agents to ground against; announced as shared between CoWork and Snowflake CoCo (the agent for coders), so the company's two agent products now sit on one grounding substrate. (2) Multi-agent orchestration: CoWork breaks complex requests into specialised sub-agents coordinated behind the scenes, with User Skills + a Skill Catalog to discover, share and reuse workflows. (3) MCP- native integrations to Gmail, Jira, Slack, Salesforce, Microsoft 365 and Google Workspace — the warehouse becomes the agent's grounded planner and the SaaS the actuator. Two reads. (1) Pair with yesterday's pattern (Microsoft + xAI + NVIDIA bringing the model in-house): the same week the model layer comes home, the data-platform layer puts its own agent surface in front of every CIO with a Snowflake seat licence. (2) Cortex Sense is the cleanest vendor-side answer this quarter to the "agents hallucinate against my warehouse" objection — and it's an MCP-server-shaped answer, which means the July 28 stateless RC lands into a unified surface across the three biggest data platforms.

The first cross-platform agent that takes the data- model problem seriously. On June 4, Sweep — the agentic-layer-for-enterprise- systems startup — announced a cross-platform agent that reasons simultaneously across Salesforce, Snowflake and Data 360 through a continuously updated dependency graph: agents can see how enterprise data, workflows, metadata, permissions and business logic connect before making any change. The pitch lands the cleanest version of an idea other vendors are dancing around — a coding agent that edits a CRM field but doesn't know the field is fed by a Snowflake view is the single most common cause of production breakage on an enterprise agent rollout. Sweep frames itself as the governance graph the agent traverses before it acts. Two reads. (1) Pair with Snowflake CoWork (item 05) and Anthropic's brake-pedal memo (item 01): the dependency-graph layer is exactly the substrate an enterprise needs in place if an agent is going to be allowed to make autonomous changes without a human in the loop on every step. (2) This is also the shape of the next M&A target list: whichever cloud — AWS, Azure, GCP — is least comfortable with Anthropic's adoption lead will be looking for a Sweep-shaped acquisition to close the "agents that know what they're about to break" gap.

The infrastructure-ops side of the same playbook. On June 1, Itential announced the general availability of FlowAI at Cisco Live US 2026: a production environment for designing, deploying and running AI agents on enterprise network and infrastructure, with the governance, security and audit primitives baked into the Itential Platform. The architecture is three pieces. FlowAgents — task-oriented reasoning agents that pursue goals through governed workflows, with full reasoning traces preserved for audit. FlowAgent Builder — the application surface for building role-based agents with defined purposes, toolsets and policy boundaries. FlowMCP Gateway — extending Itential's authentication and policy enforcement to external infrastructure agents and MCP tools. FlowAI is generally available from July 1, 2026, after six months of validation across telecom, financial services and utilities through the FlowAI Innovation Program. Two reads. (1) "Agent that pages someone at 3 AM" is the use case enterprise infra teams have been waiting to see governed properly; FlowAgents' preserved reasoning traces is the audit primitive the regulated buyers in that segment have been asking for. (2) Pair with item 05 and item 06: the same week, three separate vendors — Snowflake (data), Sweep (SaaS), Itential (infra) — converge on the same answer: the agent's job isn't to be clever, it's to be auditable.

The cleanest validation yet for the "text-only AI is a local maximum" thesis. On June 4, Generalist AI — the Pete-Florence-founded (ex-DeepMind RT-2, PaLM-E) robotics foundation-model startup based in San Mateo, two years old — announced a $400M round at a $2B valuation led by Radical Ventures, with 8VC, Union Square Ventures, Norwest, Hanabi Capital joining and existing backers NVIDIA and Bezos Expeditions following on. The company released GEN-1 in April — a general-purpose foundation model for robotics that, per its own report, hits 99% task success on benchmarks where prior models scored 64%, completes dexterous tasks roughly 3× faster, and needs only an hour of robot data per task. Use of funds: scaling models, real-world data collection, compute, and commercial deployments. Two reads. (1) Pair with Apoha (last edition, Liquid State Intelligence) and Brian Chesky's rumoured design-first lab: every funded "non-chat-box" AI primitive this quarter is in a different modality, and the capital is moving accordingly. (2) NVIDIA on the cap table for the second time in two weeks (Kumo last edition, now Generalist) signals what Jensen Huang has been telegraphing on calls — the physical-AI stack is the next platform NVIDIA wants to own, not just supply silicon to.

The week's most genuinely contrarian raise. On June 4, Flourish — a New York neuro-AI lab co-founded by neuroscientist Thomas Reardon (CTRL-labs) and former Amazon S-team executive Rob Williams — closed $500M in total funding at a reported $2.5B valuation. Jeff Bezos initially committed roughly $50M and nearly doubled his stake after other named investors — Lux Capital, GV (Alphabet's venture arm), Catalio Capital — piled in. The pitch: a system called Cortex AI that targets the human brain's energy envelope — roughly 20 watts, vs the 30×-higher draw of a single chip inside a modern AI training cluster — by reverse-engineering what Reardon and Williams call the brain's core learning algorithm. The two problems framed: power efficiency and continuous learning, both arguments that the scaling-only thesis runs out before AGI. Two reads. (1) Read with Anthropic's brake-pedal memo (item 01): if the next capability inflection comes from architectures rather than scale, Anthropic's regulatory ask is less of a brake and more of a hedge — and the capital is hedging too. (2) Bezos doubling down inside the round is the first time a single LP has publicly increased a check after seeing the cap table since the Series H era of OpenAI / Anthropic — a marker that the "watt budget" narrative is being underwritten by very specific capital, not just bench science.

The agentic-market-intelligence layer prints the kind of fundamentals the funding market is now rewarding. On June 3, AlphaSense — the AI-enabled market-intelligence and workflow- orchestration platform — announced a $350M round at a $7.5B valuation (nearly 2× the prior $4B mark) led by Vitruvian Partners with Accenture Ventures, J.P. Morgan Asset Management, D. E. Shaw Ventures, Pinegrove Opportunity Partners, CapitalG, Goldman Sachs Alternatives and Viking Global participating. The disclosure under the headline: AlphaSense crossed $600M ARR in Q1 2026, up from $500M reported in October 2025. Cumulative funding now sits over $1B. Samantha Greenberg takes CFO; Sophie Bower-Straziota from Vitruvian joins the board. Two reads. (1) The growth shape — $500M to $600M in roughly six months at a roughly $7.5B mark — is the cleanest agent-vertical comp this quarter and is the implied yardstick the next horizontal agent platform raise will be priced against. (2) Pair with Anthropic's enterprise lead (item 02) and Sweep (item 06): the analyst-workflow vertical is where the "agent reads structured + unstructured data and produces an opinion" use case lives, and the first $7.5B comp tells you exactly where the strategics start writing checks.

The headline OSS agent on the watch list — last covered as the v0.15.2 release that crossed 175k stars — clears its next threshold. On June 2–3, Nous Research released Hermes Desktop, a native cross-platform front end for Hermes Agent v0.15.2, in public preview under the MIT licence. The build: macOS, Windows and Linux desktop clients that share the same agent core, configuration, API keys, sessions, skills and memory as the CLI and gateway — i.e. not a separate product, but a new surface on the same runtime. The headline UX bit is streaming tool output — Hermes shows tool calls and their stdout as they happen, rather than waiting for the next message boundary, which is the missing piece on most terminal-shaped agents. Two reads. (1) Hermes is the first major open agent to ship a graphical surface that's not an Electron-wrapped IDE, and that distinction matters: the agent is the product, the surface is incidental — exactly the framing the closed labs avoid because it commodities the chat box. (2) Pair with item 12: the open agent layer is shipping a complete vertical (runtime + skills + memory + desktop UI) on the same week JetBrains is putting a one-year stability badge on Koog. The OSS leg of the agent stack is consolidating its product surface in real time.

The enterprise-Java side of the agent stack picks up its first credible long-term framework. At KotlinConf '26 on May 28, JetBrains released Koog 1.0 — its JVM (Kotlin + Java) framework for building production AI agents across backend services, Android, iOS, JVM, and in-browser environments. The disclosure that matters to enterprise architects: Koog 1.0 ships with a one-year API stability guarantee on stable modules — the first explicit "no breaking changes" commitment from a major-vendor agent framework. The 1.0 line splits modules into stable and beta streams, redesigns the Java interop layer, decouples HTTP transport from Ktor, brings OpenTelemetry to Kotlin Multiplatform, and adds Anthropic prompt caching out of the box. Two reads. (1) The stability guarantee is the line a JVM shop's architecture-review board has been waiting for — agent frameworks have churned hard, and "no breaking changes for a year" is the only kind of commitment that survives a regulated-bank procurement cycle. (2) Pair with item 11: between Hermes Desktop's graphical-but-not-IDE surface and Koog's JVM-native API stability, the open / vendor- neutral agent layer is shipping the exact two things the closed labs charge for — a UX shell and a stability contract.

Two MCP-shaped primitives the agent stack now has a canonical answer for. ChromeDevTools/chrome- devtools-mcp — the official Chrome team's "Chrome DevTools for coding agents" project, trending across June with broad enterprise adoption — exposes the full CDP surface (network inspection, performance traces, DOM snapshots, breakpoints) as MCP tools, replacing the hand-rolled browser-automation harness every coding agent has been shipping. D4Vinci/Scrapling — an adaptive Python scraping framework — pairs an anti-bot-resistant scraper with a first-class MCP server mode, so an agent doesn't need to know about Playwright, headers, or fingerprint spoofing to read the open web reliably. Two reads. (1) Both projects make a category obsolete: every vendor that shipped "browser automation for agents" or "scraping helper for LLMs" as a proprietary feature now competes with a single MCP face that any harness can call. (2) Pair with the MCP 2026-07-28 stateless RC (covered earlier): horizontal MCP servers like these — pure-function tools with no session state — are the cleanest early validators of the stateless protocol; the browser and the open web are now in the protocol's first-class set.