CHAPTERS
- 2:16 – 3:22
Welcome to Tokyo + Launch of Claude 5th-Gen Models (Mythos 5 & Fable 5)
Caitlin Les opens the first Code with Claude event in Japan and announces the release of Anthropic’s fifth-generation Claude models. She frames the event as a look at models, platform, and developer products aimed at expanding what builders can ship.
- •First Code with Claude event hosted in Japan
- •Announcement: Claude Mythos 5 and Claude Fable 5 released hours earlier
- •Positioning: most capable Claude models to date
- •Event theme: models, platform, and products for developers
- 3:22 – 4:53
Developer Impact Stories: Rakuten’s Agent Teams & Canva’s No-Code Mini Apps
Caitlin highlights how customers are already shipping real-world value using Claude, especially in Japan and APAC. Examples show Claude moving beyond copilots into agent-driven workflows and end-user creation.
- •Rakuten: moved from Claude Code to Managed Agents across departments
- •Agent “team lead” pattern coordinating multiple internal agents
- •Faster shipping cadence: biweekly releases vs quarterly
- •Canva: “Canva Code” lets non-coders generate interactive mini apps inside designs
- 4:53 – 5:25
How Fast Capability Has Moved: From Commit Messages to Long-Running Agents
The keynote traces the rapid evolution of model capability over the last couple of years, emphasizing shorter intervals between bigger jumps. Caitlin uses milestones in coding autonomy and security research to illustrate the accelerating curve.
- •Early frontier: drafting simple commit messages
- •Opus 4 era: model could build an entire feature
- •Agents: overnight autonomous task completion
- •Mythos: found a 27-year-old OpenBSD vulnerability
- 5:25 – 6:56
The “Capability Gap” + Why Anthropic Builds a Platform
Caitlin argues that while model intelligence is improving exponentially, most businesses realize benefits only linearly—creating a gap. The Claude platform is presented as the mechanism for developers to close that gap with scalable agents and tooling.
- •Fable 5 benchmark leadership across domains
- •Gap: AI capability vs real-world adoption/outcomes
- •Platform leverage: developers create more value than the vendor alone
- •API volume growth: ~17x year-over-year
- 6:56 – 9:23
What’s New This Week: Managed Agents Scheduling + Vaulted Secrets + Claude Code Updates
Caitlin previews the day’s agenda and announces new platform features shipping immediately. The focus is on making agent deployment easier, more secure, and more productive for developers.
- •Agenda: models (Diane), platform/agents (Angela & Caitlin), Claude Code (Kat)
- •Managed Agents: scheduled deployments (run agents on a cadence)
- •Managed Agents: vaults for secure environment variables/keys
- •Claude Code: new agent view and dynamic workflows; heavy weekly usage by developers
- 9:23 – 10:54
Diane Penn: Introducing Fable 5 & Mythos 5 and the Exponential Value Thesis
Diane recaps the rapid release cadence of Claude versions and introduces Fable 5 and Mythos 5 as the first fifth-gen models. She explains Anthropic’s belief that as intelligence rises, the value of possible use cases grows exponentially.
- •Diane’s history shipping many Claude versions across families
- •Fable 5: most capable generally available model; Mythos 5 shares foundations
- •Models already accelerating internal Anthropic work
- •Claim: agentic coding is exponentially more valuable than earlier autocomplete
- 10:54 – 12:26
Why Fable 5 Wins at Coding: Single-Shot Correctness + Long-Horizon Autonomy
Diane explains where Fable 5 differentiates most strongly: complex tasks completed correctly on the first attempt and the ability to stay coherent for extremely long runs. She emphasizes multi-day autonomy, large-token coherence, and sub-agent management.
- •Top performance on Suitebench Pro; bigger lead on longer/harder tasks
- •Single-shot correctness on complex, well-specified problems
- •Long-horizon autonomy: can run for days on a single goal
- •Handles millions of tokens; dispatches and manages sub-agents cost-consciously
- 12:26 – 13:27
Beyond Coding: End-to-End Knowledge Work + Best-in-Class Vision Understanding
Fable 5 is positioned as a workflow-scale model that can handle messy, multi-threaded tasks across common business artifacts. Diane also calls out a major jump in vision capability for technical visuals and UI-like content.
- •Handles documents, slides, spreadsheets, and financial analysis end-to-end
- •Professional-grade outputs; strong instruction-following and scope control
- •Excels on ambiguous/messy, multi-threaded requests
- •Improved vision: reads technical images, charts, diagrams, and web apps accurately
- 13:27 – 16:31
Safety Trade-offs: Glasswing, Domain Routing, and Mythos 5 Access
Diane addresses the dual-use nature of higher intelligence, especially in cybersecurity and life sciences. She describes a safeguard system that routes sensitive-domain requests to a different model and explains Mythos 5 availability via Project Glasswing.
- •Project Glasswing introduced due to potential misuse risk
- •Sensitive topics (cyber/bio/chem) route to Opus 4.8 with labeling and Opus pricing
- •Known issue: legitimate researchers may be blocked; system still improving
- •Mythos 5: same base as Fable 5 but safeguards lifted; available to Glasswing partners; expanding researcher access
- 16:31 – 21:01
What Developers Should Build Now: Time Horizon, Proactive Ownership, and Upgrade-Ready Architectures
Diane reframes capability in terms of “time horizon”—how long an agent can work before losing coherence—and argues Fable 5 enables more proactive, ownership-style agents. She closes with practical guidance: design for future models, keep harnesses simple, prototype aggressively, and treat upgrades as opportunities.
- •Metric: time horizon for autonomous coherence
- •Shift from task execution to goal ownership (weekly project tracking; living forecasts)
- •Design for the next model version; keep primitives simple as intelligence rises
- •Use harder evals/prototypes to detect newly-working experiences
- •Make upgrades easy with automated evals and repeatable testing
- 21:01 – 22:02
Angela Jiang: What an AI-Native Company Looks Like (Harness, Context, Infrastructure)
Angela opens with an AI-native vignette—software that notices issues, fixes itself, and deploys changes overnight. She outlines the three ingredients required to turn model intelligence into business outcomes and positions Claude Managed Agents as the integrated solution.
- •AI-native workflow: detect, diagnose, fix, deploy, and document without tickets/standups
- •Three necessities: harness, context, infrastructure
- •Claude platform supplies: agent harnesses, context management, production-grade infra
- •Managed Agents + Fable 5 pairing reduces effort while improving long-running agent outcomes
- 22:02 – 24:28
Managed Agents Deep Dive: Separating ‘Brain’ and ‘Hands,’ Memory/Skills, and Reliable Scale
Angela explains how Managed Agents operationalize agent work via sandboxes, iteration toward outcomes, and scaling fleets for reliability. She highlights large context, persistent memory, skill writing, and self-improvement (“dreaming”) as key context primitives.
- •Harness: tools, environment, permissions—AI that acts, not just suggests
- •Brain vs hands: model decides; sandboxes execute; iterative outcome-seeking loop
- •Context: 1M context window plus memory for continuity
- •Agents can write/read skills to fill knowledge gaps
- •Infrastructure: autoscaling sandboxes and fleets for persistent long-running agents
- 24:28 – 27:02
Customer Examples + Demo Setup: Notion/Asana + Shinkiro Racing Dashboard
Angela cites customer implementations to show Managed Agents embedded into real products and workflows. Caitlin returns to demonstrate a fictional racing team dashboard backed by multiple managed agents with rubric-driven outcomes and observability.
- •Notion: agent orchestration inside workspace for long-running delegation
- •Asana: AI teammates collaborating inside projects
- •Demo narrative: racing optimization as an AI-native workflow
- •Dashboard backed by multiple managed agents (aero, tires, power unit, safety) with ‘Outcomes’ rubrics
- 27:02 – 30:10
Caitlin Demo: Observability, Scheduled Runs, Memory + Dreaming (Shipping Today)
Caitlin walks through the developer console experience: inspecting sessions, tool calls, and results, then scheduling a recurring deployment. She shows how Memory captures learnings during runs and Dreaming retroactively improves skills/memory across past sessions—features available immediately.
- •Developer console as control plane for productivity and observability
- •Rich session logs: responses, tool calls, run completion status
- •New feature: schedule deployments (e.g., nightly safety check)
- •Memory: write findings/decisions to filesystem for future runs
- •Dreaming: review prior sessions to update memory and skills for better performance
- 30:10 – 34:37
Kat Wu: Claude Code’s Mission + New Interfaces for ‘Multi-Clauding’
Kat positions Claude Code as the bridge from idea to shipped product, evolving alongside increasing model capability. She reviews how Claude Code expanded from CLI to IDE and added new control-plane experiences to manage many parallel agent sessions.
- •Mission: shrink distance between idea and shipped product
- •Shift in workflow: from manual review of every edit to auto mode with PR-ready outputs
- •Surfaces: CLI for power users; IDE to follow code changes
- •New interfaces: Claude Code on Claude Desktop + Agents View in CLI
- •Built on Claude Agent SDK (used by extensions and many developers/enterprises)
- 34:37 – 37:40
Claude Code at Scale: Review Agents, Mobile Control, Routines, Security Scanning, Dynamic Workflows
Kat lists product primitives built from user feedback, spanning review automation, running from phone, scheduled/triggered routines, security scanning, and large parallelized work execution. She emphasizes composability as the strategy for adapting to the future of engineering.
- •Code review product: multi-agent PR bug catching
- •Remote Control + iOS/Android for coding on the go
- •Routines: scheduled/webhook/API-triggered Claude Code runs
- •Claude Security: overnight scans + vulnerability triage into Claude Code sessions
- •Dynamic Workflows: deterministic parallelism across tens/hundreds of agents for refactors/migrations
- 37:40 – 41:14
Live Demo: Localizing a Marketing Site with Dynamic Workflows (13 Languages in Parallel)
Kat demonstrates translating a website first with a single agent, then scaling via a dynamic workflow that runs many translation agents simultaneously and verifies outputs. The demo highlights how one prompt can coordinate a large, repeatable, saveable workflow.
- •Baseline: single-agent Japanese localization takes ~3 minutes
- •Problem: sequentially translating 12 more languages would take ~1 hour
- •Dynamic Workflow: create repeatable translation process and run agents in parallel
- •Desktop side pane shows 12 translation agents, then verification agents
- •Workflow can be saved as JavaScript and reused
- 41:14 – 42:24
Closing: Three Layers (Models, Platform, Code) and the Call to Explore
Kat ties the keynote together: frontier model capability, platform infrastructure for production agents, and developer tools for shipping faster. She invites attendees to join research talks, platform sessions, and Claude Code workshops—all running on Fable 5, available today.
- •Unifying narrative: capability curve + managed infrastructure + developer leverage tools
- •Dynamic workflows, Claude Desktop app, and Fable 5 available now to users
- •Encouragement to explore tracks: research, platform, Claude Code workshops
- •Theme: closing the gap between possible and deployed outcomes
