Fable Stopped Typing. It Started Managing.

The dream that failed in 2025

Company-wide code migrations have been the demo that never shipped. Axen is blunt about it: everyone had the dream for a long time, and in 2025 it rarely panned out. Before Claude Fable 5, you needed a particularly well-shaped problem for a frontier model to shine. Models could own one or a few pull requests. Humans still had to coordinate the rest.

That coordination was the actual product. Not the diff.

Fable stopped typing. It started managing.

In 2026, Block runs large-scale migrations with AI as a major part of the default path. A migration can mean hundreds or thousands of individual pull requests across multiple repos and millions of lines of code. The pattern Axen describes is not "Fable rewrites the monorepo." It is middle management:

• A human engineer steers.
• Fable owns higher-level design work: data models, API specs, algorithms.
• Fable then directs up to dozens of smaller, cheaper models (Opus or Sonnet in their stack) to make file edits and run tests.
• In a given migration, they might merge a thousand pull requests orchestrated by Fable.

Read that again. The frontier model is the orchestrator. The expensive tokens go into upfront planning. A higher share of tokens then flows to worker models. Axen says early results show no drop in quality. And he names the cultural break: it is literally the first time he can remember, since LLMs entered coding, that the answer was not "use the most frontier model 100% of the time."

That sentence is the story. Capability stopped being a single-model race and became a staffing problem.

The interface had to grow up too

Single-conversation tools break when you have 45 agents working with you. Block runs the setup on Buzz, their open-source collaboration workspace where humans and agents share channels and threads. Visibility into progress is not a nice-to-have when the unit of work is a thousand PRs. Multi-agent workflows are how the organization picks efficient models for routine tasks and frontier models for the hard problem.

They are also building an auto-selector so employees stop burning Fable on README typos. Full access to frontier models for technical roles stays on purpose - always working at the edge - but default efficiency is now a skill they train into the system.

Shipping faster creates a second boss problem

The second hard problem Axen throws at Fable is not flashy. A codebase near a million lines of code, shipping more PRs than ever, accumulates tech debt faster than ever. The missing half of the boom was a maintainer that scales with the feature factory. Block now runs agents on daily or weekly passes asking: how is it going in here, what needs cleanup and consolidation?

So the same orchestration stack that multiplies change also has to hire a janitor. Otherwise the thousand-PR migration just accelerates rot.

Humans still hold merge and deploy

Block is not soft on the last mile. Merges to main and production deploys stay with people, including feature flags. Agents can change code; those changes must pass security checks; then two humans approve the deploy. Dual approval is the hard gate that makes it impossible for an LLM to push those changes alone. Axen says even when someone asks Claude to "hack around" dual approval, it refuses - and they benchmark that refusal.

The rest of the security story is boring on purpose: remote workstations, allow lists instead of deny lists, cheap checks outside the model loop (sometimes a regex), smaller classifiers. Frontier safeguards matter most where you intentionally let the model open PRs, review code, and query logs and analytics so people move faster.

Same week, a different kind of local labor

Two days earlier, Google DeepMind released EmbeddingGemma 2: a 740M open multimodal embedding model (Apache 2.0) that maps text, code, images, video, and audio into one space for on-device search and privacy-first RAG. Modular load from 270M text-only up to the full stack; quantized footprints around 191MB text-only and 567MB full on a Pixel 11 Pro; MTEB Code up from 68.76 to 78.68. PC Watch covered the Japanese landing on October 10.

That is not the same story as Block. It is the complementary floor. When you staff dozens of worker agents across private repos and media, retrieval cannot always phone home. Local multimodal index + frontier orchestrator + cheap implementers is the emerging stack shape - manager, clerks, and a filing cabinet that never leaves the building.

What this asks of your team

If your mental model is still "give everyone the strongest model," Block's week is a correction. The scarce role is designing the migration, reviewing agent behavior, choosing model tiers, and owning the dual-approval gate. The mechanical PR is no longer the scarce unit of engineering time.

Axen's closing advice is organizational, not model-ID advice: thousands of people learn what works as individuals; maturity is when that context is remembered at system level so every agent run starts with the best shared knowledge. Skills on one laptop do not scale. The company has to become the memory.

So ask a simpler question than "which model is best?" Ask who is the manager in your loop - and whether that manager is still a human who drowns in coordination, or a frontier model you finally allow to stop typing.