CHAPTERS
Welcome to Code with Claude Tokyo + launch of Claude 5th-gen models
Caitlin Les opens the first Code with Claude event in Japan and announces the just-released fifth-generation Claude models: Claude Mythos 5 and Claude Fable 5. She frames the keynote around models, platform, and products, and why enabling developers is Anthropic’s highest-leverage path to impact.
Customer momentum: Japan-led agent adoption and new app experiences
Caitlin highlights real-world examples showing how Claude is changing shipping velocity and unlocking no-code creation. She spotlights Rakuten’s internal agent rollout and Canva’s interactive mini-app creation directly inside designs.
The capability curve: from commit messages to long-running agents and security breakthroughs
The keynote traces rapid jumps in model capability over shrinking time intervals. Caitlin cites landmark milestones—feature-building autonomy, overnight agents, and Mythos finding a decades-old vulnerability—to illustrate accelerating progress.
The “gap” problem: AI capability growing faster than business adoption
Caitlin argues most organizations still realize only linear gains despite exponential model improvements. She positions the Claude platform as the mechanism to close the gap by making it easier to build and operate powerful agents.
Platform announcements: scaling agents with scheduling + secret management
Caitlin previews new Claude Managed Agents capabilities designed for production deployments. Two headline additions are scheduled deployments and secure vault-based environment variables for authenticated tool use without exposing keys.
Diane Penn: introducing Fable 5 and Mythos 5—why they’re different
Diane explains the fifth-gen lineup, describing Fable 5 as the most capable generally available model, with Mythos 5 sharing the same base. She frames “the exponential” as intelligence increases yielding exponentially larger use-case value.
Fable 5 in practice: single-shot correctness, long-horizon autonomy, and code reading
Diane details why Fable 5 excels on complex engineering tasks: it gets more right on the first attempt and can stay coherent over extremely long runs. She emphasizes it’s not only strong at writing code but notably better at reading, debugging, and improving existing systems.
Beyond coding: end-to-end knowledge work + stronger vision understanding
Fable 5 is positioned as a workflow-scale model for messy, real organizational work—documents, slides, spreadsheets, and multi-threaded requests. Diane also highlights a major jump in vision: reading technical images, diagrams, plots, and web apps more accurately.
Safety tradeoffs: safeguard routing for sensitive domains + Mythos access via Glasswing
Diane explains that higher intelligence increases misuse risk, especially in cyber/bio/chem. Fable 5 uses a safeguard system that routes sensitive requests to a different model with clear labeling and pricing, while Mythos 5 (with safeguards lifted) is limited to Project Glasswing partners and select researchers.
Developer playbook for the exponential: design for upgrades, simpler primitives, stronger evals
Diane closes with guidance for building products that keep pace with rapid model improvement. She argues teams should architect for the next model, prefer simpler primitives that smarter models can leverage, and treat upgrades as recurring business opportunities backed by robust evaluation and testing.
Angela Jiang: what an AI-native company requires—harness, context, infrastructure
Angela describes an AI-native operating model where systems detect issues, fix themselves, and ship changes without human ceremonies. She introduces Claude Managed Agents as the packaged solution combining harnesses, context management, and production-grade infrastructure—especially powerful when paired with Fable 5.
Managed Agents deep dive: brain vs hands, memory/skills/dreaming, and reliable scaling
Angela breaks down how Managed Agents operationalize agentic work: separating decision-making from execution, enabling iteration toward explicit outcomes, and supporting deep context via large windows and persistent memory. She emphasizes the infrastructure layer—automatic sandbox lifecycle and fleets—to make long-running agents reliable at scale, citing Notion and Asana deployments.
Caitlin demo: Shinkiro Racing dashboard + new Managed Agents features in action
Caitlin demonstrates a fictional F1-style racing dashboard powered by multiple Managed Agents focused on aerodynamics, tire temps, power unit, and driver safety. She shows developer tooling and observability in the console, plus new capabilities: scheduling runs, memory, and dreaming to improve future performance.
Kat Wu: Claude Code evolution—interfaces, enterprise adoption, and developer productivity
Kat positions Claude Code as the bridge between an idea and shipped software, evolving from careful human review to delegated “auto mode.” She outlines Claude Code’s surfaces (CLI, IDE, Desktop, Agents View) and the underlying Agent SDK, highlighting broad enterprise adoption and major productivity gains.
Claude Code feature suite: review agents, mobile remote control, routines, security, dynamic workflows
Kat lists major product primitives shipped in response to user feedback, aimed at reducing review load, enabling coding anywhere, automating recurring work, and improving security and large-scale change management. Dynamic Workflows is highlighted as the orchestration layer for parallel, deterministic multi-agent execution.
Live workflow demo: website localization to 13 markets with parallel agents
Kat demonstrates translating a marketing site: first with a single agent translating to Japanese, then scaling to 12 additional languages via a Dynamic Workflow. The workflow runs translation agents in parallel, spawns verification agents afterward, and can be saved as reusable JavaScript for repeatable operations.
Closing synthesis: three layers—model capability, managed agent infrastructure, developer shipping velocity
Kat ties the keynote together: Diane’s model capability curve, Angela’s platform for controlled agent infrastructure, and Claude Code’s developer-facing acceleration. The keynote ends with a call to explore research talks, platform sessions, and workshops—all powered by Claude Fable 5, available today.
