Forge Field Report · Insider briefing · Keynote debrief · July 2026
Code with Claude: Tokyo, and the Fable 5 moment
What Anthropic actually told the room, stripped of the keynote theatre and rebuilt for builders, creators, and operators, not just developers.
The techniques, the strategy, the trade-offs they skipped over, and what to do on Monday.
The 60-second version
- Anthropic shipped its fifth generation: Claude Fable 5 (generally available) and Claude Mythos 5 (same model, safeguards lifted, partners only). Fable 5 is now the default for agentic work.
- The real message was strategic, not a spec sheet: model capability is on an exponential, your business is on a linear. The gap between the two is the entire opportunity.
- The winning move repeated three times by three speakers: build for the next model, not the current one. Keep your scaffolding thin and let intelligence do more.
- Agents got real primitives: outcome rubrics, a memory file system, “dreaming” (self-review), scheduled runs, and secret vaults. These are not dev-only ideas. They map straight onto content, ops, and research.
- Claude Code's headline shift: from reviewing every edit to delegating and reviewing the pull request. Plus fan-out across hundreds of agents in one prompt.
- The catch nobody dwelt on: a safeguard reroute that bills sensitive topics at premium prices and occasionally blocks legitimate work, plus the lock-in question that comes with all of it.
The headline: two models, one foundation
Diane Penn, Head of Product Management for Research. Diane has worked on every version of Claude since Claude 2, 21 versions across Haiku, Sonnet, Opus, and now Fable and Mythos.
A few hours before the keynote, Anthropic released the fifth generation of Claude. Two models, built on the same foundation. Claude Fable 5 is the most capable model they have ever made generally available. Claude Mythos 5 is the same underlying model with the cyber and bio safeguards lifted, available only to Project Glasswing partners.
What Fable 5 is actually good at
- Single-shot correctness. Give it a complex, well-specified problem and it tends to nail the first pass. Early testers reported single prompts producing work that would have taken teams days or weeks.
- Long-horizon autonomy. It can run for days on one goal and stay coherent, holding your specifications across millions of tokens, and dispatch sub-agents while keeping them on track and cost-conscious.
- Reading beats writing. It is better at reading code than writing it: tracing outages, digging through repo history to find what broke and when, and surfacing improvements unprompted.
- Knowledge work end to end. Financial analysis, documents, slides, spreadsheets. It stays on scope and returns professional-grade output, and it handles messy, multi-threaded requests where the next step is not obvious.
- Best-in-class vision. Dense technical images, web apps, plots, diagrams, and charts, read more accurately than any prior Claude.
When a request touches cyber security, biology, or chemistry, Fable 5 routes it to Opus 4.8 instead. The response is labelled and billed at Opus prices. Anthropic admits legitimate researchers will sometimes hit a block or a reroute. This is the price of shipping the capability to everyone today rather than months from now.
Mythos 5 is the version with those safeguards removed. It stays inside Project Glasswing for now, with life-sciences researchers being enrolled later, on the logic that the capabilities that make biology risky are the same ones that do the most good.
The strategic thesis
Diane Penn, Head of Product Management for Research. · Opening framing from Caitlyn Les, Head of Engineering for the Cloud Platform.
The most useful part of the morning was not the model. It was the operating philosophy underneath it, stated plainly.
“Model capabilities are improving on an exponential, but most business capabilities are still on a linear. There is a growing gap between what AI can do and what it is actually doing for people.”
That gap is the thesis of the entire event. Closing it is the work. The metric Diane offered for tracking it is time horizon: how long a model can work autonomously before it loses coherence on what to do next. As that number grows, the kind of responsibility you can hand off grows with it, from “write the update” to “keep the project on track all week.”
The four operating principles
- Design for the next version of Claude, not the current one. The teams that win are the ones whose architecture and product experience can absorb the next jump in intelligence without a rebuild.
- Keep the harness thin. As models get smarter they do better with basic primitives like a file system or a sandbox, and worse wrapped in sophisticated, over-complex scaffolding. Stop building cleverness the model now supplies itself.
- Write evals for things that do not work yet. Build prototype tests for experiences that currently fail. The moment a previously-broken prototype starts passing is your signal that the ground moved, and your cue to ship something you could not ship before.
- Treat every model upgrade as a business opportunity, and make upgrades cheap. Automated evals, real testing processes, and staying hands-on with each new version is how you find out fast whether new capability opens the door to a new product.
Building agents that actually do the work
Angela Jang, Head of Product for the Claude Platform. · Live demo by Caitlyn Les, Head of Engineering for the Cloud Platform.
Angela opened with a scene: overnight, a product noticed it was broken, read its own error reports, wrote the fix, and rolled it out. By the time the team woke up, the change log was already written. No standup, no ticket. Her point: an AI-native company is not one where people use AI to do their work, it is one where the work runs on AI and people decide what the outcome should be.
She named three ingredients that turn raw intelligence into outcomes. This is the most portable framework from the whole event.
1. The harness, separate the brain from the hands
The harness is what lets a model act: tools, an environment, and permission to do something. The design pattern is a split. The brain decides what to do. Sandboxes, the hands, get spun up to execute. You do not want AI that gives helpful suggestions, you want AI that makes the change. The harness also operates on an outcome: you define what good looks like with a rubric, and the agent iterates until it gets there.
2. Context, give the model something to work with and remember
- A 1M-token context window, so agents consume large amounts of material without degrading.
- Memory. Agents write findings and decisions to a file system mid-run, then read them back on future runs to do a better job.
- Self-authored skills. Agents can read and write their own skills to fill knowledge gaps.
- “Dreaming.” A separate pass where the agent reviews all its past sessions and rewrites its own memory and skills. Memory is the real-time write, dreaming is the retrospective consolidation.
3. Infrastructure, the part nobody wants to build
Long-running autonomous agents need scale and reliability that is genuinely hard to get right. Claude managed agents spin sandboxes up and down automatically and can generate multiple agent fleets on demand, which is why partners reported building agentic systems roughly 10x faster than rolling their own harness, context, and infrastructure.
A fictional racing team's dashboard ran four research projects (aerodynamics, tyre temperature, power unit, driver safety), each backed by a managed agent driven by an outcome rubric. Live, they scheduled a nightly safety check for 9am and hit a “dream” button to make the agents review past sessions and upgrade themselves. New today on the platform: scheduled deployments and secret vaults (so agents make authenticated calls without ever holding the keys).
Claude Code, in practice
Cat Woo, Head of Product for Claude Code. The average developer using it now spends around 20 hours a week with Claude.
If Section 3 was agents you ship to customers, this is agents that ship work for you. The headline is a behaviour change, not a feature.
A year ago, the workflow was: give Claude Code a task, then review every single edit and correct it on every small detail. Now many users run auto mode, delegate permissions, and only check in after Claude Code has tested its own changes and opened a pull request ready for review. The human moved from line-editor to reviewer.
The surfaces and the features
- Four surfaces: the CLI (most control), the IDE extension, the desktop app (full graphical control plane), and a new agents view inside the CLI. The last two exist to make running many Claudes at once, “multi-claude'ing”, manageable: see what is running, what is done, what needs you, jump in and out without losing your place.
- Code review agents that catch critical bugs in your PRs. Used across thousands of companies, including all internal Anthropic teams.
- Remote control on iOS and Android, so you can kick off and steer coding tasks away from your desk.
- Routines. Configure a task once and trigger it on a schedule, a webhook, or an API call. The manual “kick it off” step disappears.
- Claude Security, an overnight codebase scan that flags vulnerabilities by severity and lets you launch a session to fix each one.
- Dynamic workflows. Fan one job out across tens or hundreds of agents in a deterministic structure, for refactors, migrations, audits, and performance passes.
The localization demo translated a marketing site into 12 languages at once. One agent per language ran in parallel, then a second wave of verifier agents checked the work, and the whole thing was saved as reusable JavaScript. The shape, fan out a fleet, then fan out a verification fleet, then save the workflow, generalizes far beyond translation.
The part they aimed at developers, translated for everyone else
Synthesis. The keynote framed these as engineering primitives. They are not. Here is what each one means once you step outside the IDE.
Almost every technique on stage is domain-agnostic. A rubric does not care whether it is grading code or a newsletter. A memory file does not care whether it stores API decisions or your brand voice. This table is the bridge.
| Technique | What Anthropic showed | What it means outside engineering |
|---|---|---|
| Outcome + rubric | Define “good,” let the agent iterate until it hits the bar | Write the rubric for a lead magnet, a podcast show note, a forecast, or a sales email, and let the agent self-correct to it instead of you re-prompting |
| Memory file system | Agent writes findings to disk, reads them back next run | A persistent voice, brand, and “decisions made” store every content or research agent reads before it starts, so it stops re-learning who you are |
| Dreaming | Agent reviews all past sessions, rewrites its own playbook | A scheduled retrospective: the agent reads last month's posts, trades, or ops runs and updates its own checklist. Compounding improvement without you in the loop |
| Dynamic workflows / fan-out | 12 translations in parallel, then a verifier wave | One post into 13 platform-native variants in one action. One report into every client's localized version. The verifier wave becomes your QA pass |
| Scheduled runs / routines | Nightly safety check, webhook and API triggers | Nightly content drafts, weekly pipeline reviews, recurring research digests, fired on a cron or a webhook with no human start |
| Brain vs hands | Reasoning model separate from execution sandboxes | The same split as a coordinator-plus-voice stack, or a router in front of cheap workers. Keep the thinking expensive and rare, the doing cheap and parallel |
| Thin harness | Smarter models do more with fewer primitives | Stop over-engineering prompts and tooling. Hand the model a file system and clear goals, and delete the scaffolding it has outgrown |
| Time horizon | How long an agent stays coherent on one goal | Measure it on your own agents. It tells you exactly which jobs you can hand off whole, and which still need a human checkpoint |
Action steps
What to actually do with this, split by who you are. Pick your lane, or steal from all three.
→ If you build software
- Move one real workflow to auto mode this week. Review the PR, not the edits, and note where you still felt the need to intervene. That friction is your eval list.
- Take one ambitious-but-boring job (a migration, a codebase audit, a batch of translations) and run it as a dynamic workflow with a verifier wave. Save it. You now own a reusable asset, not a one-off.
- Stand up an automated eval harness before the next model drops, so a Fable 5 to Fable 5.x upgrade is a config change, not a project.
→ If you run a business or operations
- Pick one process that runs on a cadence (a weekly review, a month-end pack, a research digest) and rebuild it as a scheduled agent with an outcome rubric. Define “good” once, in writing.
- Find your own time horizon number. Run an agent on a real multi-step task and see where it drifts. That boundary is your delegation map for the next quarter.
- Start a memory file for the business: decisions, constraints, voice, do-not-do list. Every agent reads it first. This is the cheapest leverage on this entire page.
→ If you create content or media
- Build the fan-out pattern for distribution: one core idea, then parallel agents producing platform-native versions, then a verifier pass against your voice rubric.
- Use dreaming as a monthly ritual. Point an agent at everything you published, let it find what landed and what did not, and have it rewrite its own brief. Your style guide stops being a static document.
- Keep the brain and hands separate: an expensive model for ideas and judgment, cheaper parallel workers for drafting and formatting. Spend intelligence where taste lives.
Things to consider
The keynote was a sales pitch as much as a briefing. These are the trade-offs that did not get stage time, and they matter more the more you commit.
- The reroute tax is real. Anything brushing cyber, bio, or chemistry gets bumped to Opus 4.8, labelled, and billed at Opus prices. If your work lives anywhere near security research or life sciences, budget for both the cost and the false blocks. Test your actual prompts before you assume they pass.
- The exponential cuts both ways. “Design for the next model” also means anything you hard-build to today's ceiling may be dead weight next release. Do not over-fit your product to a limitation that disappears in three months. Build the prototype eval for what does not work yet.
- Lock-in is the unspoken cost. Managed agents, memory, dreaming, vaults, and dynamic workflows are genuinely strong, and they are Anthropic-specific. The more outcomes you wire through them, the harder portability gets. Decide deliberately which layers you are willing to make proprietary and which you keep model-agnostic.
- “Available today” is not “production-hardened.” Keynote demos are clean and scoped. Your data is messy and your edge cases are real. The demo proves the shape works, not that it works on your repo or your workflow. Your own evals are the only thing that tells you the truth.
- The bottleneck moves to verification and taste. When you can fan out to hundreds of agents, generating work stops being the constraint. Judging it becomes the constraint. The scarce input is increasingly a human who knows what good looks like, which is exactly why the rubric is the asset, not the agent.
Three speakers, one story: capability is climbing an exponential, the gap to real-world use is the opportunity, and the people who win are the ones who keep their scaffolding thin, write rubrics instead of instructions, and build for the model that ships next, not the one in front of them.
Source & method
This report is a synthesis of the Code with Claude keynote in Tokyo, drawn from the session transcript. Speaker names and titles are credited as given on stage. Company and product names have been normalized from an automatic transcript, so minor spellings may differ from official sources. Figures (17x volume, 8x code, 90% migration time, and similar) are quoted as stated by Anthropic speakers and have not been independently verified. Prepared as an internal field briefing for builders and creators.