Jonathon's AI Wiki

Code with Claude 2026 Tokyo — Fable-5-Day Re-Run of the London Breakouts

4 min read

Source: raw/Code_with_Claude_2026_Tokyo.md (YouTube GiqyYQdYoIY, Anthropic’s “Code with Claude 2026 Tokyo” developer conference, fetched 2026-06-10) — auto-caption transcript, normalized per claude-ai/AGENTS.md. Speaker names + customer metrics are caption-derived; spot-verify before external citation.

Anthropic’s first Code with Claude in Japan, held the same day as the Mythos 5 launch. The substance re-ran the London engineering breakouts (mostly different speakers, same talks), so most of it is already in the wiki. This page records the event and routes to the canonical London articles, then captures the handful of genuine deltas — it deliberately does not re-explain content covered elsewhere.

What re-ran from London (pointers, not re-coverage)

  • Cost & context engineering (Tokyo: Brad Abrams & Rod Howworth) → the same 5-step stack — prompt caching (90% off, off rate-limit) → Tool Search Tool → Programmatic Tool Calling → Compaction → Advisor strategy — is the canonical Getting more out of the Claude Platform (Puneet Shah).
  • The Capability Curve (Tokyo: Theo Chu) → The Capability Curve (Jeremy) — three capability axes + four adoption patterns (build evals, shrink scaffolding, give the model room, close the loop).
  • Managed Agents in production (Tokyo: Jess Yen & Michael Cohen) → How to get to production faster with Managed Agents — agent/environment/session model + the user/agent/session/span event taxonomy.

What’s genuinely net-new at Tokyo

  • Canva AI 2.0 (Danny Wu) — a production agentic-system case study with no London equivalent. Written up standalone: Canva AI 2.0 — Building a Production Agentic System with Claude (success-criteria design, disposable-harness/durable-evals, cache-preserving model routing, feedback-to-evals loop).
  • Cost-talk deltas over Puneet Shah’s London version: a 6th lever — model fallback + WIF (list backup models in the Messages API so a Fable safety-classifier block auto-falls-back; Workload Identity Federation removes stored API keys) — plus the new console cache-hit-rate dashboard and the tip that the built-in claude-api skill will add cache_control headers if you say “improve my cache hit rate.”
  • Managed Agents mechanism deltas: the vault secret-injection mechanism (opaque placeholder token in the container; real secret injected only at request time) and dreaming (agent reflects on past runs → curated memories; Rakuten reported ~90% of initial agent mistakes addressed). Folded into Managed Agents § Tokyo update.
  • Timing: the keynote (Caitlyn Les / Diane Penn / Angela Jen / Cat Woo) framed the day around the Fable 5 launch “a few hours ago” — Tokyo is the model’s launch-day developer event.

Key Takeaways

  • Tokyo 2026 is a re-run of the London breakouts on Fable-5 launch day — treat the London cluster as canonical for the cost playbook, capability curve, and managed-agents production model.
  • The genuine adds are Canva AI 2.0 (standalone case study), model fallback + WIF, the console cache-hit dashboard / claude-api skill tip, and the managed-agents secret-injection mechanism + dreaming.
  • When the same talk re-runs at a new event, the wiki records the event and links the canon rather than re-articling the content.

Open Questions

  • Did dynamic workflows go GA? The keynote says dynamic workflows “just went GA this week,” but the article still frames it as a research preview. Single auto-caption line — reconcile against the official changelog before flipping the status. ^[ambiguous]
  • Speaker names + customer metrics (Spotify 1,000+ PRs/mo & 90% migration cut, Rakuten 90% mistake reduction, etc.) are auto-caption-derived / stated in-talk — verify against an official recording before external citation.

Graph View

Backlinks