Anthropic completed its public strategic identity shift from "model company" to "agent platform company" at this conference, at the cost of accepting the Colossus 1 compute agreement—signing with Musk, who called them "evil" three months prior.
Counter-consensus insight: The most important signal from this conference is not any single feature release, but Anthropic's first public admission that "the bottleneck is not model capability, but compute supply"—when 80× growth pressure forces it to accept conditional compute dependency, the brand capital of the safety-first narrative is being partially converted into scale-execution capital.
The core of this conference was not any single feature, but Anthropic's public strategic identity shift. The background data is staggering: annualized revenue from end of 2025 $9B → $30B on the eve of the conference (3.3× growth); Q1 2026 actual growth 80×, far exceeding the internal forecast of 10×; API call volume YoY 17×, tokens YoY 70×.
This "happy crisis" forced Anthropic to publicly admit: the bottleneck is no longer model capability, but compute supply and workflow architecture. CPO Ami Vora's "no new models" declaration was a deliberate narrative inversion, shifting the competitive plane from model scores to agent runtimes and enterprise contract density.
A Lead Agent decomposes tasks and delegates to sub-agents, each with its own model, prompt, tools, and MCP servers. All agents run in the same container and file system, with results aggregated back into the Lead Agent's global context.
Conceptually overlaps with LangGraph and CrewAI, but Anthropic has made it a managed service with constraints and observability. The 1-level depth limit is a key design choice—enforcing flatness over the uncontrolled risk of deep recursion. Customer case: Netflix platform team uses it to analyze hundreds of build logs in parallel.
Paradigm shift: from "imperative prompting" to "rubric-defined evaluation"—developers write markdown rubrics describing "what success looks like," and an independent Grader Agent evaluates output in a fully isolated context window, preventing reasoning path contamination.
Default max iterations 3, upper limit 20. Anthropic internal testing showed up to 10 percentage point improvement (.docx generation +8.4%, .pptx +10.1%). Medical document review company Wisedocs reduced review time by 50%.
Essentially a "productized Ralph loop"—upgrading the community's hand-rolled "loop→judge→retry" pattern to a first-class API citizen, making agent work verifiable, rather than "run it and hope for the best."
Essence: an asynchronous, scheduled offline introspection process that does not modify model weights. It scans the existing Memory Store + up to 100 historical sessions, distills patterns, cleans duplicate entries, and generates plain-text notes and "playbooks."
Supported models: claude-opus-4-7 and claude-sonnet-4-6. Requires dual beta headers: managed-agents-2026-04-01 + dreaming-2026-04-21.
Customer case: legal AI company Harvey saw task completion rate increase approximately 6×—the most cited customer data point at the conference.
Dreaming's real value: providing a "experience consolidation" middle layer for long-horizon agents, sitting between RAG and fine-tuning. It doesn't learn new capabilities, only organizes old notes—media packaging it as "AI learning from mistakes" is over-marketing. Security warning: Pluto Security noted that prompt-injected "false memories" can persist across sessions via the Memory Store.
Boris Cherny's assessment hit the nail on the head: "Going forward, a lot of code will be written asynchronously." Claude Code's product boundary expanded at this conference from a CLI tool to a continuously running layer spanning CLI, IDE, desktop GUI, Web, mobile, CI integration, and scheduled tasks.
| New Feature | Status | Core Value |
|---|---|---|
| Routines (Async Automation) | Research Preview | Packages Claude Code configuration for execution on Anthropic cloud, supports scheduled / API / GitHub triggers, no local machine needed |
| Code Review | Research Preview | Dedicated agent analyzes PRs for logic errors, security vulnerabilities, edge cases, annotating with line-level inline comments |
| CI Auto-fix | Released | PR owners "will never see a red X"—Claude automatically pushes fixes when CI fails |
| Remote Agents | Released | Control Claude Code processes on your laptop from your phone, file system stays local |
| Desktop App | Released | Upgraded from CLI to a full-screen GUI that can run multiple sessions simultaneously |
Rate limits doubled (direct benefit): Pro/Max/Team/Enterprise five-hour rolling window doubled; Pro/Max peak-hour downgrades eliminated; Opus API Tier 1 input tokens/min +1500%, output +900%.
⚠️ Weekly limits not increased—explicitly confirmed by Anthropic engineering lead Amol Avasare.
Subverts the traditional "large model coordinates small model" approach: Sonnet 4.6 or Haiku 4.5 drives the agent loop end-to-end; Opus 4.6/4.7 is invoked at key judgment points, reads the full shared context, and returns plan/correction/stop signals (only 400–700 tokens), without calling tools or directly facing the user. Enable via: anthropic-beta: advisor-tool-2026-03-01.
Why Colossus 1: It is the only compute commitment among all of Anthropic's options that can deliver capacity within May 2026—Amazon's 5GW will only reach about 1GW this year; Google/Broadcom's 5GW won't arrive until 2027. After SpaceX merged with xAI, Colossus 1's GPU utilization was only about 11%, with training already migrated to Colossus 2, leaving overall capacity available for rent.
Attached clause (important): Musk retains the right to revoke compute if "Claude engages in actions harmful to humanity," but the criteria for judgment are self-determined by Musk. Simon Willison commented: "This is a novel type of supply chain risk for Anthropic."
Environmental controversy: Colossus 1 has been powered by 30+ trailer-sized natural gas turbines for nearly the past year, without Clean Air Act permits or pollution control equipment, evading regulation by classifying as "temporary use." Memphis has been designated the "asthma capital" of the US; the NAACP issued an official statement condemning Anthropic's disregard for the health of historic Black communities.
Ami Vora's keynote opening case: Mythos discovered a 27-year-old TCP SACK implementation vulnerability in the OpenBSD source tree (a signed integer overflow bug added in 1998, enabling remote DOS of any OpenBSD host responding to TCP). Approximately 1000 scaffold runs, total cost < $20,000, with the specific run finding the bug costing < $50.
Concurrent discoveries: FreeBSD CVE-2026-4747 17-year remote code execution vulnerability, FFmpeg 16-year H.264 codec vulnerability (undiscovered by 5 million fuzz tests, located by Mythos through direct code analysis).
Project Glasswing Alliance: 12 founding partners (including AWS, Apple, Microsoft, Google, CrowdStrike, Palo Alto Networks), with Anthropic committing up to $100M in usage credits + $4M in donations to open-source security organizations.
| Old Paradigm | New Paradigm |
|---|---|
| Single prompt → single response | Routines configuration → continuously running cloud agents |
| Developers write toolchain harness | Anthropic hosts agent runtime |
| Memory cleared at session end | Memory Store + Dreaming accumulate across sessions |
| Humans evaluate output quality | Grader Agent auto-scores, execution agent self-iterates |
| Single model execution | Lead Agent + specialized sub-agent fleet |
The engineering essence of this shift: Claude Code's competitors are no longer just Cursor or Copilot, but the agent scaffolding that enterprise IT departments build in-house. Anthropic packages "sandbox + state + tools + MCP + memory + multi-agent scheduling" into standardized SaaS, making self-built solutions no longer cost-competitive.
This architecture is essentially a middle layer between RAG and fine-tuning—it doesn't modify weights (avoiding catastrophic forgetting risks from continuous learning), but achieves cross-session experience accumulation through curated memory stores.
Potential security risk: prompt-injected "false memories" can persist across sessions via the Memory Store, an attack surface not present in RAG systems.
MCP has been donated to the Agentic AI Foundation (AAIF, under the Linux Foundation), co-founded by Anthropic, Block, and OpenAI. Install base 97 million (March 2026), past the network effect tipping point.
2026 roadmap priorities: stateless HTTP transport (horizontal scaling) / Tasks primitive (async long-running tasks) / MCP Apps (standardized UI components) / centralized registry. Official SDKs: TypeScript / Python / C# / Java / Swift / Go / Rust.
This is a replay of the Kubernetes and CNCF playbook: Anthropic standardizes the interface, then competes on the managed implementation. Donating MCP to a neutral foundation ensures competitors adopt the same standard—while Anthropic's Managed Agents win the premium as the most complete implementation.
| Supplier | Scale | Chips | Characteristics |
|---|---|---|---|
| AWS (Amazon) | $100B over ten years | Trainium2/3 | Primary training cloud, Project Rainier |
| Google Cloud | 5 GW, $200B over 5 years | TPU (from 2027) | Accounts for 40%+ of Google Cloud backlog |
| Microsoft Azure + Nvidia | $30B Azure credits | Nvidia GPU | M365 distribution bundling |
| SpaceX/Colossus 1 | 300 MW, 220K GPUs | H100/H200/GB200 | Available within a month, primarily inference |
Claude is the only frontier model simultaneously available on AWS Bedrock, Google Vertex AI, and Microsoft Azure Foundry—three-way distribution, not locked into any single one.
SF (5/6) + London (5/19) + Tokyo (6/10), each with an Extended Day for independent developers, globally livestreamed for free—this is Anthropic's strategic positioning of its version of "WWDC."
The real compounding is long-term: JetBrains 2026 survey shows Claude Code enterprise adoption grew 6× YoY, the fastest growth rate among AI coding tools, which is precisely the early payoff of sustained ecosystem building.
OpenAI: No counter-announcements targeting Code with Claude during May 5–10 (Sam Altman was testifying in court). But April–May saw consecutive releases of GPT-5.5, GPT-5.5 Instant, and Codex for Chrome, precisely timed around Code with Claude. The SpaceX agreement largely deflates the "capacity-constrained" narrative.
Google: The most obvious counter-scheduling—on May 6, the same day, upgraded Jules (async coding agent) to global public beta, no waitlist. The real response window is May 19 at Google I/O 2026, which coincides exactly with the Code with Claude London stop.
xAI/SpaceX: Transformed from "adversary" to "compute landlord." Musk called Anthropic "evil" three months ago, then leased all of Colossus 1 on May 6—driven by Colossus 1's ~11% GPU utilization needing monetization, and SpaceX's need for a marquee customer for the "AI infrastructure revenue" story ahead of its June IPO.
Claude Code: async long-horizon agents, terminal-native, deeply agentic
Cursor: editor experience and multi-model routing, UX moat
Copilot: enterprise standardization and procurement, Microsoft ecosystem
Codex: async cloud sandbox and PR review gating
Claude Code's "favorite" rate is 5× that of Copilot, reflecting a massive gap in deep-usage developer satisfaction. Market size $12.8B (2026) → forecast $30B (2032), growth rate supporting stratified survival for all four.
| Dimension | US (Anthropic-led) | China (Multi-vendor dispersed) |
|---|---|---|
| Foundation Models | Claude Opus 4.7 | GLM-5.1 / Qwen 3.x / DeepSeek V4 |
| Dev Tools | Claude Code + Cursor + Copilot | AutoClaw / ArkClaw / Bailian Plan |
| Runtime | Managed Agents (hosted) | Open-source framework-led |
| Compute | AWS + Google + SpaceX multi-cloud | Huawei Ascend + domestic clouds (regulation-driven) |
| Distribution Channels | Microsoft 365 + three major clouds | DingTalk + Feishu + WeCom + Xinchuang ecosystem |
| Market Strategy | Closed platform, high value, vertical depth | Open weights, low price, Global South penetration |
Manus Incident (2026/4/27): China's NDRC formally blocked Meta's $2–3B acquisition of Manus, adding agentic AI / multi-step task execution to the controlled technology catalog—this is a landmark event in China's strategic regulation of agentic AI, signaling that cross-border agentic AI acquisitions and expansion will face ongoing scrutiny.
Google I/O 2026 (5/19) will counter-attack head-on
The date coincides exactly with the Code with Claude London stop. Gemini 4 release is highly probable; the next-gen Jules (codename Jitro) will focus on "KPI-driven / outcome-oriented" development, directly targeting Outcomes. The tagline "Manually prompting your agents is so… 2025" has already previewed the attack vector.
Colossus 1 capacity coming online eases but doesn't fully solve compute issues
Weekly limit issues persist. Watch Memphis environmental litigation progress—if it escalates to a state-level injunction, it becomes an Anthropic operational risk. Musk's "Humanity Clause" is an uncertainty requiring ongoing monitoring.
Microsoft Build 2026 (6/2–3) is Microsoft's response window
GitHub Copilot multi-model strategy will be reinforced; the channel competition between Claude M365 add-in and Copilot will become public.
First real customer data for Dreaming / Outcomes / Routines
Harvey's 6× task completion rate is the only existing reference point. Mercado Libre's Q3 2026 "90% autonomous coding" target is the most important observation window. If Dreaming security risks are publicly verified by researchers, it will trigger governance discussions.
Claude Code next major version release window approaching
Current Sonnet 4.6 reaches 80.8% on SWE-bench Verified; Opus 4.7 Vals AI financial benchmark 64.37%. The next major version is expected to focus on computer use reliability (Vercept team integration results) + longer continuous task support.
SpaceX IPO (expected June–August, valuation $1.75–1.8 trillion)
Colossus 1 → Anthropic contract is the core "AI infrastructure revenue" story for SpaceX's S-1. A successful IPO will further stabilize the Anthropic-SpaceX relationship. Orbital compute (multi-gigawatt satellite data centers) will become an important component of the IPO narrative.
① Compute scarcity will permanently change competitive logic
As agent workflows proliferate, model inference demand will far exceed all current capacity forecasts. Companies that control compute allocation (cloud giants, SpaceX, chip vendors) will possess unprecedented negotiating leverage. Anthropic's multi-cloud, multi-supplier system is the most pragmatic hedge against this risk.
② Managed Agents is the most important moat, and also the biggest bet
Success or failure depends on two assumptions: enterprises are willing to outsource agent runtimes to AI vendors; Anthropic's managed service can sustainably lead on reliability, security, and observability. If true—Anthropic becomes "the AWS of the agent era"; if false—it regresses to a token pipeline. Mercado Libre Q3 2026 is the most important early validation point.
③ The end state of the "AI platform wars" is ecosystem alliances, not a single winner
MCP multi-party adoption, Claude simultaneously on three major clouds, Codex running as a plugin within Claude Code—these facts all point to: the AI coding market will not see an Office Suite-style winner-take-all. Anthropic's strategic success depends on whether it can become "the preferred agent runtime," not "the only one."
1.Compute is the binding constraint for the AI industry in 2026–2027 (not model capability)
2.Agent runtimes have moved from developer self-built into the vendor-managed competition phase
3.The competitive unit in the AI coding market has evolved from "tools" to "systems"
It changed one thing about Anthropic: the public transition from "safety-first research institution" to "scale-first platform company"—at the cost of partial dilution of brand capital, with the gain being strategic depth in compute and enterprise markets.
Code with Claude 2026 will not be remembered as the launch day of some new model, but as the inflection point where the AI platform wars shifted from "who can build the smartest model" to "who can orchestrate more electricity, chips, and workflows"—Anthropic underwrote this shift with a compute deal with Elon Musk, who was calling them evil three months earlier, writing the most dramatically tense footnote to it all.