Code with Claude SF Keynote Recap — Vora, Cherny, Jiang Anchor the Agentic-SDLC Spine; Boris Demos Web UI Catching an Edge-Case Bug
Yesterday’s opening keynote at Code with Claude SF held the agentic-SDLC line from start to finish. Chief Product Officer Ami Vora opened, Head of Claude Code Boris Cherny ran the marquee demo — Claude adding refunds to ACME’s dashboard with idempotency, multi-currency handling, and audit logging for the compliance team, working in the in-development Claude Code Web UI and discovering an edge-case bug live on stage — and Head of Claude Platform Product Angela Jiang anchored the platform track. The named customer marker that traveled the wire fastest: Mercado Libre’s 23,000-engineer org and Shopify both targeting “90% autonomous coding by Q3 this year.” Cherny’s framing on stage and on the morning CNBC hit was that the work has shifted from writing code to designing the harness around an agent that writes it.
The product spine across the keynote: Claude Code on Desktop generally available on Mac and Windows for paid plans, the “full-screen GUI for Claude Code” with drag-and-drop workspaces, integrated terminal and file editor, faster diffs, expanded previews, and SSH support on Mac. Code Review — the multi-agent PR pipeline Anthropic runs on nearly every internal PR — is now in Team and Enterprise research preview. Cloud Routines and the web/mobile entry points keep long-horizon work going while the laptop is closed. Both the IDE and Desktop surfaces are built on the same Claude Agent SDK external developers can use. Sonnet 4.8 was the watched model surface going in; the search wire as of this morning shows the Sonnet 4.8 announcement still tracking through third-party coverage rather than a confirmed Anthropic blog post, which puts the formal model line in the May 6–13 corridor we already had circled.
Dreaming, Outcomes, and Multi-Agent Orchestration Ship to Managed Agents — Memory + Dreaming Form a Self-Improving Agent Stack; Harvey Reports ~6x Completion-Rate Lift
Anthropic’s most surprising keynote drop: Managed Agents now “dream.” Dreaming is a scheduled review pass over an agent’s past sessions and memory store — it surfaces recurring mistakes, common workflows the agent converges on, and shared preferences across a team, then curates memory between runs. Memory captures what an agent learns inside a session; Dreaming refines that memory between sessions and keeps it in sync across an agent fleet. Operators choose whether dreaming updates memory automatically or routes proposed updates through a human review gate before they land. Dreaming launches in research preview. The same release lands Outcomes (you define what success looks like and Claude iterates the run to it) and Multi-Agent Orchestration (a lead agent decomposes work, dispatches sub-agents, and the Claude Console renders each sub-agent’s actions and outputs for review) — both in public beta.
Customer markers from the on-stage and post-keynote cycle: Harvey reports ~6x completion-rate lift on legal work since adopting the new orchestration; Netflix’s platform team uses multi-agent orchestration to analyze batch jobs in parallel and surface only the patterns worth acting on. Read structurally: this is the “long-horizon, low-supervision” corner of the Anthropic agent thesis put on a product surface. The Auto Mode work on the Claude Code side and the Dreaming/Outcomes work on the Managed Agents side are now the two halves of the same idea — tunable human-in-the-loop gates plus an explicit self-improvement loop — one for code, one for knowledge work.
SpaceX Colossus 1 Deal Lands — Anthropic Doubles Claude Code Rate Limits, Peak-Hour Throttling Removed on Pro and Max, Higher Opus API Limits
Anthropic and SpaceX disclosed a compute partnership giving Anthropic all of the compute capacity at the Colossus 1 data center in Memphis — ~300MW across roughly 220,000 Nvidia GPUs, just under half of xAI’s estimated 500K-GPU fleet, online inside the month. The capacity directly funds an immediate doubling of the Claude Code five-hour rate limits across Pro, Max, Team, and seat-based Enterprise plans, with peak-hour throttling removed for Pro and Max. Higher API limits land on Opus alongside. The companies also disclosed an interest in working together on “multiple gigawatts of orbital AI compute capacity” — a follow-on track that reads as an option, not a commitment, but is the first time orbital AI compute has shown up in a first-party Anthropic announcement.
Read it next to the Google line that landed Tuesday. Inside 48 hours Anthropic has put one of the largest compute floors in cloud history on the books with Google ($200B over five years, including Broadcom TPUs starting 2027) and locked in immediate, near-term GPU capacity from a competitor’s data center to clear the rate-limit complaints that piled up after the April incident cluster. The structure separates the long-horizon strategic supply (Google/Broadcom) from the short-horizon tactical capacity (SpaceX Colossus). It also drops one of the cleanest examples yet of frontier-lab compute being sold across vendor lines: Anthropic gets the GPUs, xAI gets the cash, both get to run faster.
Bloomberg: Anthropic Pivots to a Consumer Push — Claude Is #2 Free App in the US App Store, Mobile Cold-Start Cut from ~5–6s to ~1s
Bloomberg reported Thursday morning that Anthropic, originally enterprise-first, is going harder at consumers. Since late last year, employees have been pointed at how the chatbot handles personal queries — health, travel, recipes — with explicit focus on “quality, polish, and performance.” Mobile cold-start time, the Claude product team’s named target, has dropped from five or six seconds to about a second from app open to first query. The rankings line is the eye-catcher: Claude is currently the #2 free app in the US App Store, sitting between ChatGPT (#1) and Gemini (#3) — the first time a frontier-lab consumer app has held that slot.
Read for the IPO clock and the surface area. Anthropic’s model-vendor revenue mix has been overwhelmingly enterprise; the consumer line was a footnote on the S-1 narrative through Q1. The combination of Cowork on the knowledge-worker side, the M365 add-ins on the productivity layer, the SpaceX-funded rate-limit increase removing the most-cited complaint from Pro and Max users, and a #2 App Store slot resets the consumer-revenue line into a separately-defendable narrative for the prospectus. Next watch: paid-subscriber disclosure cadence and any early signal on a free-tier feature push to harvest the App Store ranking.
The Advisor Strategy Lands — Opus Coaches Sonnet or Haiku in Shared Context for Frontier-Quality Output at Executor-Tier Cost
The Advisor Strategy is now the named pricing/quality lever for production agents. Pattern: Opus 4.6 (or 4.7) generates a short plan (typically 400–700 tokens) and answers escalations on demand; Sonnet or Haiku does the bulk of the run at executor pricing. No decomposition, no orchestrator, no worker pool — just a server-side advisor tool that the executor invokes when it needs guidance. Anthropic’s published numbers: Sonnet with Opus as advisor scored +2.7 percentage points higher on SWE-bench Multilingual than Sonnet alone, while reducing per-task cost by 11.9%. The Haiku numbers are the headline: Haiku with an Opus advisor scored 41.2% on BrowseComp versus 19.7% solo — a more-than-2x lift on a hard benchmark at near-Haiku cost.
Read for the next quarter of platform conversations. Most production agent stacks are still over-spending on Opus across the entire run because routing logic is brittle. The Advisor Strategy reframes the choice as “cheap by default, expensive only when the executor flags a hard step,” which is much closer to how teams already think about senior-engineer time on a real project. Pair with the new Multi-Agent Orchestration release and the same shape repeats one layer up — a lead agent calling Opus only at decomposition and review checkpoints, with Sonnet or Haiku sub-agents executing the rest. One named customer cited “frontier model quality at 5x lower cost” on stage. The first 30-day production traces from shops adopting the pattern will be the read.
Claude Code Desktop Goes GA, Code Review Hits Team and Enterprise Research Preview, Cloud Routines Run with the Laptop Closed
The Claude Code surfaces shipped earlier in April are now the operational spine the keynote points at. Claude Code on Desktop is generally available on Mac and Windows for all paid plans — full-screen GUI, drag-and-drop session sidebar, integrated terminal and file editor, faster diffs, expanded previews, SSH support on Mac, with the Claude Agent SDK underneath. Code Review dispatches a multi-agent harness on every pull request to catch bugs the way Anthropic runs it on nearly every internal PR — available in research preview for Team and Enterprise plans. Cloud Routines run on Anthropic’s web infrastructure (4 vCPUs, 16GB RAM, 30GB disk against a cloned copy of a GitHub repo) on schedules, API calls, or events — the named patterns are nightly Linear-bug triage, automatic SDK ports between language clients, and bespoke PR review checklists.
Pinning advice for shops walking out of the keynote: turn on Code Review on a single repo first to calibrate against your own PR baseline; scope the first Routine to a low-stakes recurring job (release-notes drafting, dependency-update PRs, weekly tech-debt summary); leave Auto Mode in explicit-confirm for production changes for at least the first sprint. Auto Mode + Cloud Routines + Multi-Agent Orchestration on Managed Agents is the full Anthropic agent stack as of this morning, and the rate-limit doubling makes the per-developer cost ceiling materially less of a blocker than it was on Tuesday.
Higher Opus API Rate Limits Land Alongside the Colossus Capacity — Bedrock and Vertex Tiers Hold Status-Page Clean Going Into Extended
The same announcement that doubled Claude Code rate limits also raised the API ceiling on Opus, the most-pinched surface for shops running the advisor pattern at scale. The Rate Limits API (live since April 25) lets administrators query the new ceilings programmatically — which is the right primitive for the rolling-7-day budget conversation that emerged after the April incident cluster. The Claude status page is clean going into Code with Claude: Extended this morning — eight consecutive incident-free days for Claude.ai, the Anthropic API, Claude Code, and the Bedrock and Vertex tiers. The April 28 78-minute multi-surface postmortem is still the next operational document on the calendar; the typical inside-ten-business-days cadence puts publication into the May 8–11 window.
Operational stack to pin for shops planning load through Friday and the weekend: Managed Agents on the harness layer with Multi-Agent Orchestration enabled selectively, Rate Limits API on the budget layer, flex tier on Bedrock and a secondary failover to Vertex on the inference layer. Pricing for Opus 4.7 still sits at $5 / $25 per MTok and Sonnet 4.6 at $3 / $15 per MTok. The new tokenizer that ships with Opus 4.7 can produce up to 35% more tokens for the same input text — reset cost telemetry baselines accordingly before drawing month-over-month comparisons.
Code with Claude: Extended Runs Today — Indie Devs and Early-Stage Founders Get the Builder Track After Yesterday’s Keynote
Code with Claude: Extended SF runs today, May 7. The format is the indie-developer and early-stage-founder day: founder stories, builder deep-dives, and laptops-open workshops led by Anthropic’s Applied AI team — no livestream, sessions recorded for post-event publication. The frame is “Code with Claude is where you hear what’s new; Extended is where you see it in the wild,” which lines up with the keynote-day product wave (Dreaming, Outcomes, Multi-Agent Orchestration, Code Review, and the Desktop GA) hitting the practitioner cohort the day after the announcement.
The Extended pattern repeats in London on May 20 and Tokyo on June 11. Read for the channel: Anthropic’s 2026 conference cadence is paired-day, three-region — one keynote day for the wire, one builder day for the product surface adoption. The same rhythm shows up in the cap-table announcements (the Wall Street JV Monday, the FS keynote Tuesday, the developer keynote Wednesday) and now in the field-marketing track. The structural read is that Anthropic is running a calendar that intentionally stacks story arcs rather than scattering them.
Mercado Libre and Shopify Target 90% Autonomous Coding by Q3 — Spotify and Epic Anchor the Cowork Story, Harvey and Netflix Anchor Managed Agents
Customer markers across the keynote and the keynote week formed a usefully complete picture of where the agent stack is being adopted. Mercado Libre’s 23,000-engineer org and Shopify both went on record with a Q3 target of “90% autonomous coding” using the Code Review and Multi-Agent harnesses. Harvey’s legal team reported a ~6x completion-rate lift on Multi-Agent Orchestration. Netflix’s platform team uses Multi-Agent Orchestration to analyze batch jobs in parallel. On the Cowork side from Tuesday’s Briefing FS keynote, Spotify reported any engineer can now “kick off a large-scale migration just by describing what they need in plain English,” and Epic noted more than half of its Claude Code usage is now from non-developer roles.
Read for the S-1 narrative. The case-study layer is no longer a thin column — it’s a stack: code automation (Mercado Libre, Shopify), agent orchestration (Harvey, Netflix), knowledge work (Spotify, Epic), and finance verticals (the Briefing FS roster). The pattern these references share is depth-of-deployment, not breadth-of-pilot, which is what enterprise procurement buyers will look for in any Q2 vendor review.
Claude Partner Network Adds ICODA — Web3 / Blockchain Marketing Joins the Roster as the Vertical-Channel Build-Out Continues
ICODA, a blockchain and web3 marketing agency, joined the Claude Partner Network this week. The Partner Network is Anthropic’s program for organizations helping enterprises deploy Claude in production — the same channel layer that is doing most of the heavy lifting on Cowork and the financial-services agent rollouts. Specialist agency announcements like this one are below the wire-news fold individually, but they mark the channel-coverage pattern Anthropic has been building behind the named system-integrator partners and the JV-funded acquisition track Reuters surfaced this week.
Read with the Reuters M&A line: both the Anthropic and OpenAI joint ventures are now in active talks to acquire AI services firms — OpenAI’s Deployment Company has three deals in advanced stages, Anthropic’s $1.5B Wall Street JV is funding tuck-ins of engineering and consulting shops. Specialist Partner Network adds (ICODA on web3, the financial-data partners landing on Tuesday, the named system integrators) plus a JV-funded acquisition track is a two-pronged channel build-out: organic for breadth, M&A for depth. Watch the next two weeks for the first named acquisition target on either side.
Wall Street Week, Day Four: The Self-Improving Agent Stack Lands on Top of a Doubled Compute Floor — the IPO Narrative Now Has a Consumer Line
Step back from the wire feed and the picture going into Code with Claude: Extended is unusually coherent for a single conference week. Inside 96 hours Anthropic has put the $200B Google Cloud and Broadcom-TPU compute floor on the books, closed the $1.5B Wall Street JV, shipped ten ready-to-run finance agents and the Microsoft 365 add-ins, run the first shared-stage moment with JPMorgan’s Jamie Dimon, anchored the “moment of danger” cyber-window narrative through Mythos and the Fed’s Bowman, taken all of the SpaceX Colossus 1 capacity to double Claude Code rate limits and remove peak-hour throttling on Pro and Max, and shipped Dreaming, Outcomes, Multi-Agent Orchestration, Code Review, the Claude Code Desktop GA, the Advisor Strategy, and the consumer push that put Claude at #2 in the US App Store. The structural shape is four layers economic-stacking on each other inside one week: long-horizon compute, vertical agents, developer platform, and consumer surface.
What changes today versus Tuesday morning: the IPO narrative now has a consumer line that is independently defendable, the agent stack has a self-improvement loop on its product surface that the bear case will have to argue against rather than around, and the per-developer cost ceiling that piled up complaints across Pro and Max is structurally lower as of this morning. The bear case still notes the same lines — concentrated cloud-vendor commitment, the Pentagon-blacklist drag, an October S-1 timeline against an active lawsuit, and the Mythos cyber-window asymmetry — but the bull case now reads with a four-layer stack on the cap-stack page, customer markers (Mercado Libre, Shopify, Harvey, Netflix, Spotify, Epic) that read as depth not breadth, and a consumer line that did not exist as a story 30 days ago. Watch the formal Sonnet 4.8 announcement (still inside the May 6–13 corridor at the time of this brief), the first 30-day production traces from shops adopting the Advisor Strategy, the April 28 postmortem inside the next four business days, and any named JV-funded acquisition target.