The agent stack gets a runtime, a regulator —
and a $200M watcher.

The Build day-two story Microsoft did not pull to the top, and the one that completes the runtime pitch yesterday's brief led with. Microsoft introduced Autopilots — a new category Microsoft defines as "always-on agents that work autonomously, with their own identity, acting on your behalf" — and shipped its first one, Microsoft Scout. Scout proactively schedules meetings across time zones, generates prep materials, blocks calendar time for upcoming deliverables, surfaces stalled decisions before they become blockers, and builds contextual understanding through Work IQ; it spans Teams, Outlook, OneDrive and SharePoint, runs across cloud / desktop / web, and is powered by OpenClaw with Microsoft contributing policy-conformance capabilities upstream. The governance model is the consequential bit: Scout runs under its own Entra identity (not a shared service account), credentials are scoped to specific tasks, redacted from logs, and access is enforced through Microsoft Purview sensitivity labels and DLP. Availability is private preview via the Frontier program; admission requires Frontier enrolment, Intune policy configuration, opt-in attestation and a GitHub Copilot license. Two reads. (1) Scout is the concrete answer to "what would a Windows-Server-for-agents look like in productivity?" — an OpenClaw harness, but with an identity model the regulated buyer can already model in their directory. (2) Microsoft picking OpenClaw (rather than commissioning a closed agent stack from a partner) is the strategic statement of the week: the open-agent runtime is now a first-class platform substrate, not a research hedge.

The IDE move Cognition has been signalling since the Windsurf acquisition closed. On June 2, Cognition shipped Devin Desktop: the same editor surface as Windsurf, but folded into a unified Devin brand spanning four products — Devin Desktop (the editor), Devin Cloud (the autonomous agent), Devin CLI, and Devin Review. Three primitives matter. (1) Agent Command Center: the default surface, a Kanban view that lets a developer manage every local and cloud agent in one place. (2) Spaces: a new container that shares context across agents while grouping sessions, PRs, files and context. (3) Agent Client Protocol (ACP): an open protocol Devin Desktop launches with native support for, so any ACP-compatible agent — Codex, Claude Agent, OpenCode, others — runs inside it. The local agent itself was rewritten: Devin Local succeeds Cascade, is built in Rust, and ships with up to ~30% better token efficiency plus subagent support. Two reads. (1) Cognition is making the same bet OpenAI did with Codex plugins yesterday and Microsoft did with VS 2026 Copilot Agents: the IDE is the agent surface, not the chat box. But Cognition's distinguishing move is the open protocol — ACP is the explicit invitation to other vendors' agents to live inside Devin. (2) The Rust rewrite + 30% token-efficiency claim is the cost-of-runs story: with credit pools live across GitHub Copilot and the Anthropic Agent SDK, the agent that uses fewer tokens wins the next procurement.

Meta's first serious enterprise-agent move, and the one most likely to land at SMB scale fastest. At Conversations 2026 in London on June 3, Meta introduced Meta Business Agent — an agentic customer-engagement tool that handles inbound on WhatsApp, Messenger and Instagram: bookings, follow-ups, product recommendations, and (Meta's framing) "close sales" — with human escalation always available. The companion Meta Business Agent Platform lets businesses build, customise and deploy their own agents and wire them into existing enterprise infrastructure. Adoption baseline at announcement: more than one million businesses already use a Meta Business Agent on WhatsApp / Messenger; the tool ships free initially and becomes part of a paid subscription tier in the coming months, with pricing scaled by business size. Two reads. (1) The distribution model is the strategic point: Meta is pointing its 3B-user messaging surface at SMB commerce agents, going directly at Salesforce Agentforce and Shopify's agent layer at the small-buyer end. (2) Pair this with Microsoft Scout (item 01): both vendors made an agent product real on June 2–3, both lean on an identity / governance model the target buyer already trusts, both bet the package-as-job pitch (covered in yesterday's Codex-roles brief) is now table-stakes.

The pieces of the Computex keynote that yesterday's brief under-printed because the OpenShell / RTX Spark / OpenClaw story crowded them out. Vera CPU is the purpose- built Arm processor for agentic / reinforcement-learning workloads — NVIDIA claims roughly 2× the efficiency and 50% faster performance than traditional x86 server CPUs, and the early-adopter list reads like the state of the lab race: OpenAI, Anthropic and SpaceX. Vera anchors the broader Vera Rubin platform — Rubin GPU, NVLink 6, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet — and is now in mass production. Alongside it, Nemotron 3 Ultra: a 550-billion-parameter MoE with 55B active that tops US open-weights at an Intelligence Index of 48, delivers 300+ output tokens/second, runs up to 5× faster inference than leading alternatives, and costs about 30% less. Underneath, NVIDIA released an updated Agent Toolkit centred on NemoClaw — an open framework for constructing agentic orchestration layers — and confirmed OpenShell Secure Runtime is co-developed with Microsoft, Canonical and Red Hat. Two reads. (1) Vera-CPU early-adopters being OpenAI / Anthropic / SpaceX is the procurement signal: the labs that publicly compete on benchmark and policy are quietly converging on the same agent-tuned silicon. (2) Nemotron 3 Ultra at Intelligence Index 48 puts NVIDIA's open-weights line credibly in the same neighbourhood as the closed frontiers — and pairs with NemoClaw to give NVIDIA the full "model + framework" stack it has to date refused to ship.

The cleanest Day-3 Computex item not already inside yesterday's brief. MediaTek CEO Dr. Rick Tsai used his keynote on June 3 to unveil two new Dimensity AX automotive parts — the C-X1, billed as the first auto chipset to run AAA-class games and agentic-AI HMI workloads on the same SoC, and the MT2739, an auto chipset pairing 5G NR-NTN satellite with edge-cloud agent collaboration. The keynote also ran a live 6G interop demo aimed at edge-cloud agent collaboration and a joint NVIDIA DGX Spark on-device agent demo. Two reads. (1) Read with NVIDIA's RTX Spark consumer PC (yesterday) and Vera CPU (item 04), MediaTek is positioning itself as the edge silicon partner to the NVIDIA agent stack — the spec-and-volume vendor that ships Vera-class primitives into automotive and edge boxes the consumer vendor doesn't cover. (2) The 6G interop demo is the tell on where the on-device agent story goes next: local first, but with a deterministic offload path the regulator can audit.

The infrastructure piece sitting under the runtime / model stories above. Vista Equity Partners and Cambium launched Vector Core Compute (VC2) — what the launch press calls the "world's first commercially-available enterprise inference cloud" for disaggregated inference. The architecture is the headline: a single workload is split across silicon classes — Intel Xeon 6 for orchestration & execution, SambaNova SN40 RDUs for decode, NVIDIA Blackwell for prefill — and a live demo ran MiniMax 2.5 across all three in one inference path. The financing line: a $3.5B compute commitment to SambaNova plus Intel support stand behind VC2 at the Los Angeles facility, with sites in development in Chicago, Seattle and Phoenix and a target of 50+ US metros. Together.ai signs on as the first commercial customer. Two reads. (1) Vector Core Compute is the operational answer to the credit-pool wave (Anthropic Agent SDK split, GitHub Copilot AI Credits): if the bill scales with tokens, the substrate that picks the cheapest silicon per phase wins economically. (2) Disaggregated inference also fixes the "supply-mix" risk hyperscalers carry — VC2 is a hedge that the next neocloud is silicon-agnostic by design, not a single-vendor footprint.

The week's largest cheque for the agent stack's operational layer. Coralogix — Boston-HQ'd software-monitoring vendor founded in Israel — closed $200M in Series F financing on June 3 at a $1.6B post-money, only 11 months after a $115M Series E and bringing total raised to $550M. Lead investors: Advent and CPPIB, with Greenfield Partners and Brighton Park Capital participating. The pitch is the moving target: Coralogix's built-in "Olly" AI agent — plus MCP and CLI surfaces for automated workflows — runs on the same telemetry foundation as its dashboards, so customers move from human-led incident response to autonomous observability without re-platforming. Growth disclosed with the round: revenue up 60%+ YoY and ~30 customers spending more than $1M / year. Two reads. (1) Observability-for- agents is the second-derivative procurement category — the budget the CIO finds after the agent is in production and the audit committee asks who authorised what — and Coralogix is the first vendor of scale to put a public price on it. (2) Read with the Salesforce Agent Fabric GA (yesterday's brief) and Anthropic's Project Glasswing expansion: governance + identity + observability are converging into one control-plane budget line, and the buyer is the same.

The consulting-channel move Anthropic spent March's $100M Partner Network launch laying groundwork for. On June 3, Anthropic announced the Services Track — a three-tier ladder for consulting and SI firms gated by certified-practitioner count, live joint customers, and public stories — and the Claude Partner Hub, a self-serve portal where partners see exactly where they stand and customers find the firms most qualified for a given scope. The thresholds: Select (≥10 certified / ≥2 customers in production over 12 months / ≥1 public story); Preferred (≥100 / ≥15 / ≥3); Global Premier (≥1,000 / ≥100 customers across 3+ regions / ≥15 stories / a joint business plan with named executive sponsors). Promotions process twice a year (Jan 1, Jul 1) with an extra Oct 1, 2026 review in this first year; standings refresh daily in the Hub. Two reads. (1) The tiering effectively gives Anthropic an enterprise go-to-market without owning the headcount — the same play Microsoft has run with the MAS / Inner Circle / GSI programs for two decades. (2) Pair with yesterday's KPMG-Anthropic 276k-employee rollout and the Bristol Myers Squibb deployment: Anthropic's enterprise wedge is not the chat UI but the certified humans behind it, and the Partner Network is the procurement vehicle for that wedge.

The China-side capital story of the week. DeepSeek — the Chinese open-weights lab that until now refused outside capital — is reported by Bloomberg / CNBC on June 3 to be closing its first-ever external funding round at roughly $7.4B and a $52–59B post-money. Lead commitments under discussion: founder Liang Wenfeng personally committing ~20B yuan, Tencent ~10B yuan, CATL ~5B yuan, with NetEase, JD.com, China's national AI fund, IDG and Monolith in talks; the round is expected to close in two weeks. Two reads. (1) DeepSeek reversing its no-VC stance is the capital-side admission that next-generation agent training (long-horizon RL, large-scale tool-use supervision) is now too capex-heavy for organic cashflow — even for the most operationally efficient lab in China. (2) ~$55B mid-range puts DeepSeek ahead of every other Chinese lab and within striking distance of Mistral / xAI — and pairs with yesterday's Anthropic S-1 filing as evidence the frontier-lab cohort is now public-market- ready on both sides of the Pacific.

The US side of the rulebook week. On June 2, the White House signed "Promoting Advanced Artificial Intelligence Innovation and Security" — a voluntary 30-day pre-release frontier- model review run via the NSA Director, a Treasury-stood-up "AI Cybersecurity Clearinghouse" within 30 days, and a 60-day framework deadline for labs (reduced from 90 in earlier drafts). On June 3, Sam Altman went to Capitol Hill, met Speaker Mike Johnson, Hakeem Jeffries and Senator Sanders, and published an OpenAI policy blueprint explicitly endorsing the voluntary-review framework; Altman also flagged a future vehicle for distributing AI windfall gains to consumers. Two reads. (1) The order is shaped to be implementable, not aspirational — the 30-day window is short enough to keep frontier ship cadence on track and long enough to give NSA / Treasury an audit hook. (2) Altman being the first lab CEO on record endorsing the framework is the political consequence yesterday's Anthropic S-1 didn't price: OpenAI is now the public co-author of the US regulator playbook, and the Anthropic-vs-OpenAI "safety lab" narrative is harder to run on regulator-relations alone.

The Brussels side of the same rulebook window. On June 3 the European Commission proposed a Tech Sovereignty Package bundling four instruments: the Cloud and AI Development Act (CADA), a Chips Act 2.0, an Open Source Strategy, and an Energy AI roadmap. CADA's headline mechanic is a four-tier "sovereignty" framework for public-sector cloud / AI procurement, gated on infrastructure location, supply-chain control and cyber posture. Reaction is split fast: the CCIA (Computer & Communications Industry Association) has already labelled it "discriminatory" and warned of "severe market fragmentation." Read with the EU's Digital Omnibus (passed May 7), which already delayed High-Risk AI Act obligations to December 2027, the direction of travel is clear. Two reads. (1) CADA is the lever Brussels will use to steer agent-platform procurement away from US hyperscalers — the tiering is sovereignty by another name, and any vendor pitching a regulated EU buyer needs a CADA-aware deployment story by Q3. (2) Pair Trump EO + Altman blueprint (item 10) with CADA: in a single 36-hour window the world's two largest agent buyers published different rulebooks. Vendors building cross-Atlantic deployments now have to ship both simultaneously.

The signal-from-floor side of the runtime story. Three Claude Code / Codex skills shipped inside the last ~10 days and are already on the trending surface: (1) app-it (Christian Katzmann, created May 30) — a Claude Code + Codex skill that turns a local project into a macOS Dock-launchable app with its own window and icon, no Electron; ~160+ stars in five days and a Windows beta in progress. (2) vibecode-pro-max-kit (withkynam, created May 27) — a spec-driven coding harness with 12 agents and 32 skills aimed at "self-improving context memory" against context-rot, ~750 stars in eight days. (3) adhd (UditAkhourii, created May 25) — a Claude / Codex Agent SDK skill that runs a tree-of-thought with pruning across divergent cognitive frames, ~730 stars in ten days. The pattern: with the in-IDE runtime fights (item 02, yesterday's VS 2026 Copilot Agents) and the package-as-job pitch (item 03, yesterday's Codex roles) both settling, the marginal builder is shipping a SKILL.md not a framework — and the agent marketplaces are starting to price discoverability. The unifying question for next week: which of the OpenClaw / NemoClaw / Claude Skills registries gets the first cross-vendor signed-skill spec, the way signed-binary distribution settled the OS layer.