CHAPTERS
- 2:01 – 3:46
Ami Vora on the joy of building—and why AI is shifting what’s possible
Ami opens with a personal story about discovering the joy of programming and contrasts the old barriers to building with today’s always-available compute and AI. She frames Claude as a “superpower” that gives more people access to that same feeling of discovery and creation.
- •Personal origin story: first program that worked and the emotional hook of building
- •How access has changed: no more lab queues or specialized infrastructure to do powerful work
- •Developers experience AI’s shift differently—frontier builders, guides, and those seeking what’s next
- •The pace of change forces constant replanning and adaptation
- •AI tools now lower barriers to creating things that didn’t previously exist
- 3:46 – 5:49
Real-world impact: faster migrations at Stripe and faster foster-care licensing at Binti
Ami grounds the “superpowers” idea in concrete examples that show both efficiency gains and real human impact. The stories highlight how AI-enabled developer work can compress timelines dramatically.
- •Stripe: 50k lines of Scala migrated to Java—10 weeks estimated, done in 4 days with Claude
- •Binti: Claude API reduces caseworker paperwork and removes 20 days from foster-family licensing
- •Efficiency as a means to outcomes—not just speed metrics
- •AI impact spans internal engineering and mission-critical social systems
- •Developers translate model capability into practical tools people actually use
- 5:49 – 7:50
The exponential capability curve vs. linear adoption—and the developer’s role in closing the gap
Ami describes rapid leaps in model capability—from email writing to long-running agents to vulnerability discovery—and argues most organizations still adopt AI slowly. She positions developers as the key bridge between frontier capabilities and real-world value.
- •Rapid capability milestones: hour-long agents → overnight agents → Mythos finding a 27-year OpenBSD vuln
- •Capability intervals are shrinking while jumps are getting bigger
- •Organizations adopt AI linearly, creating a widening capability-to-impact gap
- •Developers are the mechanism for turning capability into usable products
- •Platform growth signals: ~17X API volume YoY; ~20 hours/week average Claude Code usage
- 7:50 – 9:21
What to expect today: no new model, but major product upgrades across platform and Claude Code
Ami previews the keynote structure: model-layer updates, platform primitives for agents, and Claude Code primitives for automation. The throughline is enabling developers to ship faster and more reliably using improved tooling.
- •No new model announcement; focus is on making products work better for developers
- •Agenda: frontier model direction; managed agents updates (Outcomes, Dreaming, orchestration)
- •Infrastructure focus: platform handles scaling so teams can focus on building
- •Claude Code upgrades: routines and new primitives for self-prompting automation
- •Most users experience AI via products built by developers—not direct API calls
- 9:21 – 10:37
Rate limit and compute expansion: doubling Claude Code limits and raising Opus API limits
Ami announces expanded rate limits enabled by new compute partnerships and investments. The goal is to directly empower individual developers and small teams to build more with Claude.
- •Doubling Claude Code 5-hour rate limits for Pro/Max/Team/seat-based enterprise plans
- •Raising API limits considerably for Claude Opus
- •Compute partnership: SpaceX Colossus One data center capacity
- •Investment targeted at individual developers and small teams
- •Commitment to explore additional compute strategies and “bolder bets” over time
- 10:37 – 12:08
Diane’s model-layer recap: the evolution of Claude’s frontier capabilities
Diane reviews the pace of model releases since Claude 2 and highlights key capability breakthroughs across versions. She frames the model layer as the foundation that enables everything else announced in the keynote.
- •Since 2023: 18 Claude versions across Haiku/Sonnet/Opus/Mythos
- •Key improvements: JSON adherence, long-form code, safe computer use, exposure of overeager behaviors
- •Claude 4: thinking dials and test-time compute progress
- •Eight frontier models shipped in the last 12 months
- •Model intelligence moves the developer starting line forward
- 12:08 – 14:10
What “the exponential” really means: new capabilities, new products, new markets
Diane explains that the exponential is not just benchmark scores; it’s the invention of capabilities like tool use, long-horizon planning, and long-context learning. She argues these unlock qualitatively new product categories beyond coding assistants.
- •Exponential impact: agentic coding vs. autocomplete as a step-change
- •Capability invention: tool use, computer use, adaptive thinking, long agentic loops
- •Long context windows can teach Claude new domain knowledge
- •Broader outputs: visual design iteration, complex deliverables, ambiguous business problem-solving
- •Claude’s advantage: first to create key capabilities and time spent making them reliable
- 14:10 – 15:43
Opus 4.7 in production: benchmark wins, tooling simplification, and visual design strength
Diane shares examples of Opus 4.7 outperforming prior approaches and reducing the need for heavy scaffolding. She also highlights Claude Design as a complementary product driven by Opus 4.7’s visual “taste.”
- •AMP moved “smart mode” to Opus 4.7 due to top benchmark performance and reduced scaffolding needs
- •Rakuten: resolved ~3x more production engineering tasks on benchmarks
- •Intuit: model detects planning-stage logical faults, backtracks, and executes cleaner
- •Claude Design launch: production interfaces built with Claude Design + Claude Code
- •User preference: Claude understands full assignments and pushes back on assumptions
- 15:43 – 18:13
What’s next in research: higher judgment, near-infinite context + memory, and multi-agent coordination
Diane outlines three major focus areas that push toward more autonomous, trustworthy engineering work. She introduces “task horizon” as a way to measure autonomy duration and quality improvement over time.
- •Roadmap themes: higher judgment and better code taste for complex autonomous work
- •Context windows that feel infinite when paired with high-quality memory
- •Multi-agent coordination for goals too large for one instance
- •Task horizon: minutes (last year) → hours (now) → proactive always-on agents (future)
- •Models remain unfinished—basic stumbles and context loss still occur
- 18:13 – 20:45
Developer strategy for the next intelligence jump: evals, cheap upgrades, ambitious prototypes
Diane advises builders to architect for future model capabilities rather than optimizing only for today. She emphasizes robust evaluation, minimal scaffolding, and treating upgrades as a business opportunity.
- •Design for the next Claude version, not just the current one
- •Maintain harder evals to detect when previously impossible ideas start working
- •Build ambitious prototypes to surf capability jumps
- •Make model upgrades cheap via automated evals and simple scaffolding
- •Shift in scaffolding: from keeping models “upright” to amplifying intelligence
- 20:45 – 22:03
Claude Platform framing: outcomes are hard, scaling is harder—platform primitives aim to solve both
Angela and Caitlin introduce two common blockers for businesses adopting AI: getting consistent outcomes and shipping scalable production systems quickly. They position the Claude platform as providing tuned primitives, infrastructure, and operational controls for agentic systems.
- •Problem 1: steering models to the right outcomes requires prompt optimization, tools, and harness engineering
- •Problem 2: shipping fast while scaling in production is difficult
- •Claude platform offers model-tuned API primitives
- •Provides infrastructure for building and scaling agentic systems
- •Adds operational controls to run these systems reliably
- 22:03 – 23:47
Lower cost with high intelligence: the Advisor strategy (small executor + big advisor)
They present an agent architecture that separates execution from advising to reduce costs while maintaining quality. The approach allows cheaper models to execute while consulting a frontier model only when needed.
- •Implementation: update the tools array in the Messages API
- •Architecture: split execution from advising
- •Example: Haiku/Sonnet executes; Opus advises selectively
- •Observed result: Sonnet + Opus advising performs better and can be cheaper than Sonnet alone
- •Case study: eVE Legal reports frontier quality at ~5x lower cost
- 23:47 – 32:02
Managed Agents upgraded: multi-agent orchestration, Outcomes, and Dreaming—plus a lunar drone demo
They explain how Claude Managed Agents accelerates prototype-to-production and bundles best practices like memory. A live demo (“Lumara”) shows multi-agent coordination, rubric-based iteration with Outcomes, and Dreaming to self-learn from past runs and improve results overnight.
- •Managed Agents: production-grade harness + infrastructure; teams ship ~10x faster
- •Memory bundled out of the box; portable/owned by the customer
- •Upgrades: multi-agent orchestration, Outcomes (success rubrics + iteration), Dreaming (self-learning to memory)
- •Demo: commander/detector/navigator with independent threads and merged results
- •Dreaming generates a descent playbook; improves simulations from 4/6 to better performance after overnight learning
- 32:02 – 36:08
Claude Code vision and surfaces: CLI, IDE, and Desktop as an agent control plane
Cat reframes Claude Code as closing the gap between ideas and shipped software by exposing frontier intelligence to every developer. She walks through how usage evolved from hand-holding to auto mode, and how Desktop adds a full-screen control plane for local and remote sessions.
- •Mission: make frontier intelligence accessible and practical for builders
- •Usage evolution: many permission prompts → auto mode with PR review checkpoints
- •Surfaces: CLI (power users), IDE (follow code changes), Desktop (visual control plane + previews)
- •Desktop: remote session visibility, stuck/ready indicators, rich outputs
- •Built on Claude Agent SDK; internal impact: ~200% increase in PRs per engineer while maintaining quality
- 36:08 – 40:36
Claude Code primitives in production: review agents, remote control, Autofix, routines, and security
Cat lists the main developer pain points and the corresponding features shipped to automate them. She then shares how large organizations adopt these tools at scale, changing engineering workflows and even bringing managers back into hands-on coding.
- •Features: automated code review via agent teams; mobile remote control; Autofix for CI/comments/conflicts
- •Routines: configure once, triggered by webhooks/APIs/schedules to kick off work automatically
- •Claude Security: overnight scans + auto-remediation workflows via Claude Code
- •Shopify: org-wide adoption across engineering and non-engineering; internal tool velocity gains
- •Mercado Libre: 23k engineers; 500k+ PR reviews with oversight; 9k+ apps modernized; push toward 90% autonomous coding
- 40:36 – 47:29
Boris demo: async coding with routines and CI Autofix—verification as the enabling shift
Boris demonstrates Claude Code Desktop handling a real product feature (refunds) end-to-end, including detecting and fixing an edge-case race condition through browser verification. He then zooms out to show parallel session management, routines that turn issues into PRs automatically, and Autofix that keeps PRs green by retrying or fixing CI failures.
- •Demo task: implement refunds with idempotency, multi-currency, audit logs; verify in UI
- •Self-debugging: finds missing success toast, traces race condition, fixes and re-verifies
- •Desktop overview: manage many parallel sessions; see which need input vs. merged PRs
- •Routines as “higher-order prompts”: issues/webhooks/schedules trigger async PR creation locally or on remote compute
- •CI Autofix babysits PRs: handles review/security comments, fixes CI, rebases conflicts; reduces developer exposure to failures
