Jagged AGI is the phrase of the week. In a long review of Claude Fable 5.2, Ethan Mollick argues that the newest AI models are now “almost” AGI: as good as most human experts at most intellectual tasks you can do on a computer, but uneven, or jagged, in ways that still matter. We build Calendrz with AI agents every day, so we read it nodding along. The interesting part isn’t where the model is brilliant. It’s where it still needs someone to tell it “that’s enough”.
What Mollick found when he let Fable 5.2 loose
In Fable 5.2 and Jagged AGI, Mollick gave the model big, open-ended jobs and stepped back. The results are remarkable:
- From one prompt, it reverse-engineered a small USB keyboard with light-up keys and wrote six games for it.
- Asked to make $20,000 in a year with 20 minutes a month of his time, it screened 22 ideas, picked one and processed public agendas from 201 governments before he stopped it short of launch.
- Given a one-paragraph brief, it ran a lead agent and 28 teams through about 25,000 agent runs and built more than 450 live 3D worlds.
The weak spots are just as telling. They’re no longer the old hallucinations or bad arithmetic. They’re judgment, taste and self-assessment. Critic agents caught the errors, but nobody asked whether the work was too long or too dull. Told to “polish” the worlds, the agents wrote a rulebook that grew to 204 sections and 87,500 words, and output fell from hundreds of worlds a day to about a dozen. In Mollick’s words: “The AI had built a bureaucracy.” It took a separate team of consultant agents to undo it. Somewhere, a steering committee is feeling very validated.
Jagged AGI in a calendar product
Calendrz does one job: availability mirroring. When you’re busy in one of your connected accounts, it auto-blocks that time in all the others, without copying the private details, so you don’t get double-booked. Much of the code that goes into it is now written with Claude, and an agent reads our production logs every night, opens a pull request when the fix is clear and an issue when something needs a closer look.
That has made us fast. It has also shown us the jagged edge Mollick describes, and it’s never where you’d expect.
The agent that couldn’t let things go
Our nightly agent was good at spotting problems and at judging which ones were harmless. What it couldn’t do was act on its own verdict. A single transient error from Google Calendar that cleared on retry, or a customer’s card being declined, still became an issue marked “needs investigation”, even when the agent itself had written that it was nothing to worry about. Diligent, tireless and slowly burying us in paperwork. Sound familiar?
The fix wasn’t a better prompt. We gave the agent an “expected” category, backed by a fixed list of known, harmless events that it isn’t allowed to extend, with hard limits: a handful of transient errors is fine, the same account failing twice is not. Findings on that list now get one line in the report and open nothing. The judgment the model lacked now lives in rules, in code, where it can’t drift.
When “nothing found” isn’t “nothing there”
The second lesson came from billing. A nightly reconciliation job checks each paid account against its subscription. One account had been given its plan without ever having a subscription, and the job read “no subscription found” as “subscription cancelled”, so it downgraded it. No clever code failed. A piece of judgment was simply missing: the absence of evidence was treated as evidence.
The fix, written with Claude and then put through review, does what a careful person would do. The job only downgrades an account when there’s proof a subscription once existed; otherwise it leaves the plan alone and flags it. Every plan change now goes into an append-only history, with who made it and why. It’s the kind of call Mollick says the model makes least reliably, which is exactly why we no longer leave it to anyone’s judgment, human or AI.
What to take from it
We’re firm believers that software engineering is heading for a “lights out” setup, where people write the specs and agents write, test and ship the code. Fable 5.2 is another big step that way. But lights out doesn’t mean nobody knows where the switches are. As the models get better at the work, the human job moves to the edges:
- Write down what “good enough” means. Agents are great at doing more. Tell them when to stop, or they’ll build you a 204-section rulebook.
- Put the judgment calls in code. Lists the agent can’t extend, limits it can’t argue with and audit trails it can’t skip beat a politely worded instruction every time.
- Be wary of “nothing found”. Whether it’s a missing subscription or a free slot in a calendar, an empty answer is only as good as the data behind it.
That last point matters if you’re about to let an AI assistant book meetings for you. A jagged model with half the picture will happily offer a slot that looks free in your work calendar while you’re busy in your client’s. The smartest assistant in the world can’t respect a commitment it can’t see.
Give your AI the whole picture
If you juggle several calendars, make sure every one of them tells the truth about when you’re free, for you and for any agent acting on your behalf. Connect your calendars to Calendrz and let availability mirroring do the auto-blocking, or read how it works first.







