CHAPTERS
Welcome to London & why this event matters
The keynote opens with excitement about bringing Code with Claude to London for the first time. The speaker frames the day as a developer-focused look at how Anthropic is translating rapidly improving model capability into usable products.
From calculator hacks to AI agents: collapsing the idea-to-running gap
A personal origin story (TI-83 calculator programs and HTML for eBay listings) illustrates the joy of tinkering and immediacy. The talk contrasts that with modern software complexity—and argues AI is shrinking the distance between an idea and working code again.
Real-world impact: migrations at Spotify & foster care workflows at Binti
Two customer stories demonstrate that AI coding isn’t just convenience—it changes timelines and outcomes. Spotify uses Claude Code to automate large-scale repo migrations, while Binti uses Claude API automation to reduce paperwork time and speed foster-family licensing.
The capability curve is accelerating—but adoption is still linear
The keynote highlights fast, compounding jumps in model capability over short time intervals. It argues most organizations still adopt AI slowly, creating a widening gap between what models can do and what people actually get in production.
What’s coming today: models, managed agents, and Claude Code primitives
The agenda is laid out as a three-layer stack: frontier model progress, platform support for secure scalable agents, and developer-facing Claude Code features. The speaker stresses most users won’t touch the API directly—they’ll use products developers build.
Claude’s model evolution: key milestones from Claude 3 to Opus 4.7 & Mythos Preview
Lisa reviews major capability leaps across model generations, showing how each release unlocked new categories of tasks. The emphasis is on new behaviors (computer use, adaptive thinking, long-horizon planning) rather than benchmarks alone.
Customer proof on Opus 4.7 & expanding beyond coding
Lisa shares how Opus 4.7 performs in real deployments and why stronger model intelligence expands addressable markets. Claude is positioned as reliable across both software engineering and domain-specific workflows like legal, finance, and even genomics.
What’s next in models: judgment, effectively infinite context, multi-agent coordination
The talk previews the next frontiers: stronger autonomous engineering judgment, longer effective context for continuous work, and teams of agents collaborating on large goals. A key concept introduced is “task horizon”—how long a model can work before losing the thread.
Developer guidance: build for emerging capabilities and avoid over-scaffolding
Lisa advises developers to architect systems for the next capability jump, not just today’s constraints. As models improve, heavy scaffolding can become a limiter—so teams should invest in evals, prototypes, and streamlined primitives that scale with intelligence.
Platform challenges: getting the right outcomes and scaling to production
Caitlin and Anju frame two adoption blockers: it’s hard to achieve precise outcomes (prompting, tools, harnesses) and hard to scale quickly without losing quality. The Claude platform is presented as a set of tuned primitives, infrastructure, and operational controls.
Cost + quality with the Advisor Strategy (small executor, large advisor)
They introduce an approach that splits “execution” from “advising” so teams can lower costs while maintaining frontier quality. A smaller model runs most steps, escalating to a larger model only when needed—improving both performance and sometimes total cost.
Claude Managed Agents: shipping 10× faster with orchestration, outcomes, dreaming
Managed Agents are positioned as a production-grade harness plus infrastructure to build scalable agents quickly. Newer capabilities help teams coordinate fleets of agents, specify success criteria, and enable self-improvement from prior runs.
New security & control: self-hosted sandboxes and MCP tunnels (Counter GrowthBot demo)
A live demo shows how internal-agent workflows can stay secure and private. MCP tunnels provide secure access to behind-firewall MCP servers, while self-hosted sandboxes ensure code execution happens on the company’s own infrastructure.
Claude Code’s evolution: from CLI power users to multi-session agent management
Cat reframes Claude Code as tooling that makes frontier intelligence accessible and manageable for everyday development. The product has expanded from CLI to IDE and new multi-session interfaces (desktop and Agents view) to support “multi-Clauding.”
New Claude Code primitives: code review agents, mobile control, Autofix, routines, security
The keynote lists product capabilities built in response to user pain points: reducing review time, enabling on-the-go delegation, and removing PR babysitting. These features are presented as composable primitives that change the development workflow.
Enterprise adoption stories: Shopify and Mercado Libre’s org-wide agentic coding
Examples show Claude Code deployed across entire companies, not just individual developers. The narrative emphasizes how org-wide agent use shifts what work gets done—modernizing legacy systems and enabling even non-coding leaders to ship again.
Live Claude Code desktop demo: refunds feature, verification, routines, and CI Autofix
Boris demonstrates Claude Code completing a complex feature end-to-end, finding a UI edge case, diagnosing a race condition, fixing it, and verifying in a browser. He then zooms out to show async development via routines and PR babysitting via CI Autofix.
Closing: one story across three layers—capability is here, deployment speed is the gap
The keynote concludes by tying together the model curve, platform controls, and Claude Code workflows. The call to action is to explore the day’s sessions and build systems that can absorb the next leap in capability.
