A Review Agent on Fable 5 Found Everything I Didn't Check
Anthropic's new, strongest model — Fable 5 — became available for a week — and the first thing I did wasn't write code with it. I pointed it at reviewing my changes. It found a staggering number of bugs and drew up a fix plan itself.
By Georg Malahov
Yesterday ended in irritation. A seemingly simple task — swipe-back navigation in my dictation app, a web app where you speak your text and get it back clean — turned into my longest session on record: the agent ran for almost seven hours, launched 227 background sub-agents, and did do something, but did it crookedly. Refreshing the page on the result screen still doesn't work. A whole day of tiny, practically invisible features — and the feeling that the day was wasted.
And today I updated the Claude app, and the subscription got Anthropic's new, strongest model for a week — Fable 5. The limits are very tight, so I didn't spend it on writing code. The first thing I did was point it at review.
First — on yesterday's dictation task. Then — on all the changes I had piled into Call Copilot, my browser extension for call transcription, over the past weeks and never checked by hand. The result stunned me: an enormous number of bugs. In the model's opinion, almost nothing in there works at all.
The setup is simple. The review agent gets the code changes and goes through them like an outside reviewer who didn't write this code: it hunts for bugs, vulnerabilities, defects. But the key part is that the output isn't a list of complaints — it's a detailed, precise fix plan. I hand that plan to agents to execute, then review again — and so on in a circle. After the first such iteration, the dictation task got significantly better: practically everything the way I wanted it, the manual visual test passes. Still, after the first iteration the review found more bugs — I'm launching the second one now.
The review phase existed in my cycle before, but with this model it started working in a way you feel immediately — in what bugs it finds and how efficiently. It looks like I'll end up with this pairing: a regular model writes the code, and the strongest one does reviews and plans — possibly even pay-as-you-go, because it's worth it. If all the dirty work — finding the defects and making the plan — is done up front by the strong model, then what's left is to execute cheaply and check again.
In parallel I handed the agent a completely different kind of work: sorting Stripe invoices into my bookkeeping. Of two hundred two unassigned transactions, one hundred eighty-five were left by evening — it downloaded the documents itself, renamed them, started uploading and assigning categories. I hadn't been able to get to that task for months.
I have one week with this model. The plan for it is simple: review everything I can reach — and the second iteration on the dictation app will show whether the list of findings converges to trivia or whether this review can be spun forever.