Some coverage on the crazy Fable saga.

Reading this, I am reminded about @bgurley’s take of Dario’s *Machines of Loving Grace* essay on @theallinpod.

If you started a new project with agentic coding over the past 1–2 months using previous models such as GPT-5.5 or Claude Opus 4.8, chances are the next best course of action you can take right now is to rewrite your code. Hear me out. Since Claude Fable 5 launched, we are all quickly realising that Fable looks like a completely new tier of coding model, not just an incremental upgrade. On Cognition FrontierCode Main, a benchmark designed to test whether code is actually good enough to merge, Fable 5 scores 46.3, ahead of Claude Opus 4.8 at 34.3 and GPT-5.5 at 25.5. That is roughly 35% above the previous Claude frontier baseline. (*Refer to the benchmark graph below.) Here is why that matters for your existing project: given how much stronger the coding model is, previous code and prompts are no longer neutral context. They become baggage, old assumptions that anchor Fable to bad patterns and weak abstractions. Some of that code only worked because the previous model kept patching around its own limits. This radically changes the economics of technical debt from both directions at once. Holding on to old agentic code just got more expensive, because that code can actively drag a stronger model down. Paying the debt off also got dramatically easier, because the stronger model can now rewrite the weak parts directly. The old rule was to avoid rewrites unless absolutely necessary, because rewrites were slow, expensive, and risky. That rule was built for a world where rewrites took months. Stripe is the clearest sign that this world is ending. During early testing, Anthropic reported that Fable 5 “compressed months of engineering into days.” In a 50-million-line Ruby codebase, it performed a codebase-wide migration in a day that would otherwise have taken a whole team over two months by hand. Source: anthropic.com/news/claude-fa… If Fable can make migrations dramatically easier on 50 million lines of code, then on a 1–2-month-old agentic codebase, it becomes rational to rebuild before the bad assumptions harden. I have been vibe coding for almost a year, and spent the past month coding out @CalatheaAI , an AI curation system for myself. Under the old rules, rewriting after a month would look premature. With Fable, I think waiting is the bigger mistake the cost of rewriting has fallen that much, the opportunity cost has already flipped. Now, the bigger risk is not rewriting too early. The bigger risk is letting brittle crawler assumptions, weak freshness checks, and patched source logic harden into the foundation. One more thing. Given that Fable is out of the bag, the rest of the field will catch up. Open-source models will get sharply better through distillation, and future paid models will be radically more powerful still. Sooner or later, those capabilities will be pointed at all the agentic slop we have accumulated over the past year and a half since vibe coding became a thing. The vulnerabilities a weaker model wrote, a stronger model can find. So going forward, revisiting previous agentic code with Fable or an equivalent model is not just paying down "technical debt". It may become a security imperative, even a moral one, especially if you are building a product with actual customer data or finances. We have been curating all things AI daily, but today, @CalatheaAI is hunkering down and rebuilding with Fable. I think most of us should do that too.

👀

Tried Claude Fable 5 with this approach and it works well. The model seems better when you stop over-prescribing the steps. Give it the objective, define what done looks like, add verification criteria, then let it find the path. A lot of old Claude.md files may now be too restrictive.

It's here.

Claude Code just added nested subagents Anthropic’s Boris Cherny says Claude Code can now let agents launch other agents, with an initial nesting depth capped at 5. The goal is context management: instead of forcing one agent to carry the whole task, child agents can handle isolated workstreams in their own context. Worth testing today if you use Claude Code for agentic coding workflows.

Sharing this gem again in case the people at the back missed it: MIT Missing Semester is still one of the best resources for anyone starting out with coding with AI. Not because it teaches you how to prompt better. Because it teaches you the layer that prompts don’t replace. Shell teaches you how programs talk to each other: arguments, streams, env vars, return codes, signals, pipes. That matters when your AI agent is running commands, reading stdout, failing on paths, missing env vars, or returning a non-zero exit code. Debugging teaches you what to do after the model gives you 500 lines and something breaks: read logs, inspect stack traces, use verbose flags, isolate the failure, figure out whether the code is wrong or the environment is wrong. Git gives you the safety rail: diffs, branches, commits, rollback, history. Code quality tools give you the verification layer: formatters, linters, tests, type checkers, pre-commit hooks, CI. That’s why Missing Semester aged so well. AI made code generation easier, but it made the surrounding skills more important. Run the code, inspect the failure, track the diff, test the change, package the environment, ship the artifact. If you’re learning to vibe code, don’t just learn prompts. Learn the system the code has to survive in. missing.csail.mit.edu/

/loop used to mean Ralph-style retry loops. Now it’s starting to mean continuous orchestration via cron jobs: agents supervising agents, spawning threads, checking work, recovering state, and looping until verified. Basically, AI systems building themselves. Great read.