Skip to content
ClaudeClaude

Code with Claude London 2026: Opening Keynote

Get the latest updates from Anthropic's engineering and product leaders at the Code with Claude 2026 opening keynote in London.

May 19, 202646mWatch on YouTube ↗

CHAPTERS

  1. 0:58 – 1:30

    Claude comes to London: first international Code with Claude

    The keynote opens with a warm greeting to London and the excitement of bringing Code with Claude outside San Francisco for the first time. The speaker sets an informal, community tone (including a quick selfie) before moving into the story of why coding—and building with AI—feels magical again.

    • First Code with Claude event hosted outside San Francisco
    • Audience welcome and setting expectations for the morning
    • Community-first vibe and shared excitement
  2. 1:30 – 3:04

    Learning to code on a calculator: practical “magic” and early tinkering

    A personal origin story: learning to program on a TI-83 scientific calculator using TI-BASIC to solve math problems. The speaker describes teaching classmates and even publishing a programming guide at 13, highlighting the joy of tinkering and immediate feedback.

    • Programming on a TI-83 (and UK Casio comparison)
    • Using TI-BASIC programs as study aids
    • Sharing knowledge with classmates and early online publishing
    • The emotional hook: coding felt like magic
  3. 3:04 – 5:08

    From eBay HTML hacks to modern complexity—and the distance collapses again

    The narrative moves from playful HTML for eBay listings to the growing complexity of modern software development (tooling, configs, build systems). The speaker argues AI is collapsing the distance between idea and running software—recapturing the early “calculator feeling,” now at distributed-systems scale.

    • HTML learned to improve eBay Pokémon card listings
    • Programming’s toolchain complexity increased over time
    • AI reduces the gap from “idea” to “it runs”
    • The new magic: describing problems and getting working systems
  4. 5:08 – 6:38

    Real-world impact: Spotify migrations and Binti’s foster care outcomes

    Concrete case studies show AI’s value beyond novelty: Spotify uses Claude Code agents to automate large-scale repo migrations, and Binti uses Claude API to cut administrative overhead for caseworkers. The examples emphasize time savings and mission-critical outcomes.

    • Spotify: background agent fleet opening PRs at scale
    • Over 1,000 PRs/month merged; migration time cut by 90%+
    • Binti: hours returned to caseworkers; licensing cut by 20 days
    • Efficiency reframed as human impact
  5. 6:38 – 9:41

    The capability-adoption gap and why developers matter

    The keynote frames an industry-wide mismatch: models improve exponentially, but organizational adoption is linear. Developers are positioned as the crucial translators who turn raw capability into products people can actually use.

    • Model progress is exponential; enterprise adoption is linear
    • Growing gap between what AI can do and what it’s deployed to do
    • Developers close the gap by building usable systems
    • Momentum signals: API volume up ~17x; heavy Claude Code usage
  6. 9:41 – 12:06

    Today’s roadmap: model layer, managed agents, and Claude Code primitives

    The speaker previews the event structure across three layers: frontier models, Claude managed agents on the platform, and Claude Code features (agent view and routines). The segment reinforces that most end users experience AI through products built by attendees, not direct API calls.

    • Agenda overview: models → platform agents → Claude Code updates
    • Shift toward Claude managed agents and secure deployment at scale
    • Claude Code highlights: agent view and routines
    • Developers shape what AI feels like for everyone
  7. 12:06 – 13:36

    Claude’s model evolution: from code proficiency to end-to-end ownership

    Lisa reviews the rapid progression from Claude 3 to Opus 4.7 and Mythos Preview, emphasizing increasing autonomy, judgment, and ambiguity-handling. The message: eight frontier models shipped in 12 months, and model behavior is the foundation for everything else.

    • Timeline of model milestones (code, computer use, thinking, documents)
    • Opus 4.7/Mythos: owning outcomes end-to-end with judgment
    • Eight frontier model releases in the past year
    • Model capability as the base layer for product experiences
  8. 13:36 – 15:08

    Why incremental intelligence creates new markets: tool use, thinking, long context

    The keynote explains the “exponential” thesis: small intelligence gains unlock qualitatively new use cases (agentic coding vs autocomplete). Lisa highlights areas Anthropic pioneered—tool/computer use, adaptive thinking, long-horizon agent loops, and long context windows—now reliable enough for broad domains.

    • Agentic coding is a step-change over autocomplete
    • Capabilities: tool use, computer use, adaptive reasoning
    • Agentic loops that hold plans over hundreds of steps
    • Long context and reliability as differentiators
    • Expansion beyond coding into many verticals
  9. 15:08 – 16:39

    Customer results with Opus 4.7: coding agents, engineering tasks, and design taste

    Lisa shares customer proof points showing improved practical performance: AMP, Rakuten, and Intuit report better task resolution and self-correction during planning. She also introduces Claude Design (Labs) and notes users increasingly see Claude as a thought partner that can push back.

    • AMP: moved smart mode to Opus 4.7; simplified tooling
    • Rakuten: 3× more production engineering tasks resolved
    • Intuit: catches logical faults in planning; faster execution
    • Claude Design: production UI in one sitting; strong visual taste
    • User perception: “understands the assignment” and challenges assumptions
  10. 16:39 – 18:10

    What’s next: judgment, ‘infinite’ context, and multi-agent coordination

    Looking forward, the talk forecasts higher judgment/code taste, effectively infinite-feeling context, and coordinated multi-agent teams. The unifying metric is “task horizon”—from minutes last year to hours today, moving toward continuously running, proactive agents.

    • Higher judgment and code taste for complex autonomy
    • Context windows that support longer continuous work
    • Multi-agent coordination for goals too big for one agent
    • Task horizon as the key metric: minutes → hours → continuous
    • Vision: proactive always-on agents managing high-level goals
  11. 18:10 – 20:57

    Developer guidance: design for the next model, reduce scaffolding, upgrade via evals

    Lisa advises developers to architect for emerging capabilities rather than optimizing for today’s limitations. As models improve, heavy scaffolding can become a constraint; teams should rely on generalized primitives, invest in harder evals/prototypes, and treat model upgrades as business opportunities through automated testing.

    • Build for the next model generation, not just the current one
    • Excess scaffolding (loops/instructions/tools) can hold smarter models back
    • Use generalized primitives: filesystem, sandbox environments
    • Evals/prototypes reveal when previously-failing tasks start passing
    • Automate testing to make upgrades routine and value-creating
  12. 20:57 – 22:07

    Claude Platform: solving outcomes + scaling production agents

    Caitlin and Angela frame two blockers for businesses: getting the right outcomes is hard, and scaling quickly with quality is harder. They position the Claude platform as end-to-end support: model-tuned API primitives, agentic infrastructure, and operational controls.

    • Two core problems: outcome difficulty + production scaling
    • Prompt optimization, tool construction, harness engineering challenges
    • Platform provides primitives, infrastructure, and controls
    • Goal: ship fast without sacrificing quality
  13. 22:07 – 23:13

    Lower cost frontier quality: the advisor strategy (executor + advisor models)

    They introduce the advisor strategy: use a smaller model for execution and call a larger model only for guidance. The approach improves performance and can reduce cost, enabling freemium experiences and high-volume workloads with better ROI.

    • Split execution (small model) from advising (large model)
    • Example: Sonnet executor with Opus advisor improves results
    • Reported outcome: frontier quality at ~5× lower cost (Eve Legal)
    • Best for freemium products and high-volume agentic systems
  14. 23:13 – 24:17

    Claude Managed Agents: faster to production with orchestration, outcomes, and dreaming

    The platform story advances to Claude Managed Agents—an agent harness plus production infrastructure that enables teams to ship far faster. They cite Asana AI teammates and highlight core features like multi-agent orchestration, outcomes-driven iteration, and self-improvement via “dreaming.”

    • Managed Agents: production harness + infrastructure
    • Claimed speedup: build in days, not months
    • Asana example: AI teammates collaborating in projects
    • Key capabilities: orchestration, outcomes, dreaming/self-improve
  15. 24:17 – 31:13

    New enterprise controls: self-hosted sandboxes and MCP tunnels (Counter demo)

    Two major upgrades are announced and demonstrated via a fictional company (Counter) building a growth agent: self-hosted sandboxes for executing code on your own servers, and MCP tunnels to securely reach internal MCP servers behind firewalls. The demo shows proactive Slack updates, secure data access, PR creation, and rich observability in the console.

    • Self-hosted sandboxes: run execution on customer-controlled infra
    • MCP tunnels: secure access to internal MCP servers without public exposure
    • Demo: Growthbot in Slack calls experiment winner and opens PR
    • Console observability verifies tool calls, branches, and sandbox execution
    • First-class integrations mentioned: Daytona, Cloudflare, Vercel, Modal
  16. 31:13 – 40:08

    Claude Code product layer: interfaces, org-wide adoption, and new primitives

    Kat shifts to Claude Code, thanking early adopters and describing the mission: make frontier intelligence accessible to every builder. She covers the evolution from CLI to IDE, desktop app, and CLI Agents view—plus product primitives like code review agents, mobile remote control, Autofix, routines, and security scanning—supported by large-scale adoption examples.

    • Mission: close the gap from idea to shipped product
    • Interfaces: CLI, IDE, desktop control plane, CLI Agents view
    • Primitives shipped: code review agents, mobile remote control, Autofix, routines, Claude Security
    • Org adoption examples: Shopify and Mercado Libre scale outcomes
    • Shift toward async development and less babysitting
  17. 40:08 – 45:14

    Live demo: async coding with desktop sessions, routines, and CI Autofix

    Boris demonstrates Claude Code’s “magic” in a realistic engineering scenario (Acme Pay refunds): end-to-end implementation, browser verification, bug diagnosis, and fix validation. He then zooms out to show managing many parallel sessions, routines that trigger work from issues/webhooks/schedules, and CI Autofix that keeps PRs green by handling flakes, reviews, and rebases.

    • Refund feature demo: idempotency, multi-currency, audit logging
    • Claude verifies in-browser; finds and fixes race condition
    • Desktop app shows multiple parallel sessions and their states
    • Routines: higher-order prompts triggered by schedule/webhooks/API
    • CI Autofix: handles flaky CI, review comments, conflicts; keeps PRs green
  18. 45:14 – 46:26

    Closing: one story across layers—capability is here, speed to impact is the gap

    The keynote closes by connecting the day’s themes: model capability curves, managed agents on controllable infrastructure, and developer-facing tooling that makes async, verified shipping the norm. Attendees are encouraged to explore the remaining sessions and start building to put the exponential to work.

    • Three-layer narrative: models → platform → Claude Code workflows
    • Verification enables async development and trust
    • Remaining challenge: translating capability into real-world deployment speed
    • Call to action: explore talks/workshops and build with the ecosystem

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.