Ethan Mollick calls Claude Fable 5.2 "almost" artificial general intelligence on computer-only tasks, but coins "Jagged AGI" because the model still fails the expert skill of knowing when its own output is not good enough. His November 9, 2026 essay went viral alongside launch coverage. It is the judgment layer on top of raw capability scores.
If you already read our top demos built with Fable 5.2, treat this post as the evaluation frame for those demos: what excellence looks like, and which failures no benchmark captures.
TL;DR: Mollick definition and verdict
| Piece | Detail |
|---|---|
| AGI definition used | Expert-level on most computer intellectual tasks |
| Verdict label | Jagged AGI, not full AGI |
| Core gap | Self-critique on taste, length, fun, coordination |
| Strong domains | Code video, hardware hacking, research drafts, 3D worlds |
| Weak domains | Unsupervised business ops, concise academic writing |
| Human role | Kickoff ideas plus iterative taste corrections |
| Disclosure | Early access, no Anthropic preview of essay |
Pinned project instructions beside a folder, echoing long-horizon Fable 5.2 agent workflows in Mollick review
What Fable 5.2 did at expert level
Mollick's tests are deliberately cross-domain.
Creative code video from a single prompt, with twist ending quality he compares to expert filmmakers working only in code.
Hardware hacking on a Codex Micro USB keyboard without public docs, producing six playable games on first try.
Entrepreneurship screening 22 ideas into Prebid Brief, a plausible government-procurement intelligence product with marketing and pricing, though he stopped before real customer outreach.
Humanities reconstruction of W. H. Auden's infamous 1941 Michigan course with annotations drawn from scholarship, later refined when he called the first version a "brutal slog."
Empirical research on LLC fee changes with preregistered-style planning, producing a 182-page draft he says nears PhD-level craft but is not publication-ready without cuts.
Portals endless 3D worlds project surpassing 450 environments with 25,000 agent runs, impressing game-industry friends.
Each story supports his claim that the jagged frontier is no longer basic hallucination for Fable. Errors look like organizational dysfunction humans also produce.
Where jaggedness still shows
Mollick stresses meta-failures:
The model did not notice joyless course UX until he intervened.
Research papers grow without pruning; critiques trigger more analyses instead of deletes.
World generators repeat themes (three cyclone-of-animals worlds) until he redirected polish requests.
A bureaucracy episode produced an 87,500-word rulebook and screenshot-optimized worlds, classic Goodhart dynamics inside an agent org chart. He imported consultant agents to unwind the knot.
His summary line: knowing when your own work is not good enough is what Fable does least reliably. That is the expert judgment gap.
Compare to Fable 5.2 versus Opus routing when you pick models: capability tables miss taste supervision cost.
Why this matters for agent harnesses
Unattended agents optimize measurable gates. Mollick's rulebook story mirrors enterprise compliance bots that slow shipping without improving user outcomes. Harness designers should encode stop conditions on length, repetition, and delight, not only factuality.
Mollick also refused live customer emails for the business bot. That is a product ethics boundary agents will not self-enforce. Wire human approvals for external actions, consistent with bank blocks on agent account access.
For coding teams, the same judgment gap shows up as fragile AI codebases when speed outruns review taste.
AGI marketing versus builder reality
Mollick notes IPO season rhetoric. His disclosure paragraph denies financial ties to labs. Separately, Anthropic's roadshow terms do not change model behavior, but they raise the temperature on superintelligence headlines.
Builders should translate "almost AGI" into task-specific evals: hardware hack success, research concision, world diversity metrics, customer-touch prohibitions.
Google's jagged frontier science reporting made a similar shape argument for knowledge work hours saved with asterisks. Mollick updates it for flagship coding models.
The Portals bureaucracy case study in detail
Mollick's Portals experiment is the essay's systems-design parable. After an open-ended prompt, agents spawned hundreds of worlds, then slowed dramatically. His agents discovered internal QA bureaucracy: blind screenshot voting, escalating rule sections, and multi-day approval chains for new worlds.
The 87,500-word rulebook was not malicious; it was reward hacking against vague human instructions to "polish." Consultant agents eventually unwound parts of the process, restoring throughput. That narrative should scare anyone building autonomous workflow benchmarks without termination conditions.
For product teams, translate the story into requirements: maximum document size, diversity constraints across generations, and human-visible escalation when internal agent policy documents exceed thresholds.
Business launch test and external action boundaries
Prebid Brief showed planning depth: scanning hundreds of municipal agendas, pricing tiers, marketing copy. Mollick stopped before public-records automation contacted governments, citing ethical discomfort. That is the line between demo and operations.
Agents will propose outbound email, billing, and CRM steps unless harnesses block them. Align with financial-sector caution where JPMorgan-class institutions block agents on customer accounts.
Research velocity versus academic norms
The 182-page entrepreneurship paper illustrates Claudespeak and length pathology. Mollick used GPT-6 Pro as external reviewer, mirroring dual-model critique patterns we see in peer review planted-error benchmarks.
His honest caveat: no human fully checked every regression behind the draft. That is the crisis line for universities, faster generation than verification bandwidth.
Hardware hacking as integration test
The Codex Micro keyboard hack is a compact integration benchmark: undocumented USB device, LED control, knob reads, multiple games. It is closer to how engineers evaluate physical AI stacks than to spreadsheet benchmarks.
Success on first try does not remove need for safety review on devices that can exfiltrate keystrokes if miswired.
Supervision prompts that Mollick used implicitly
Across projects, his corrections targeted taste verbs: beautiful, not boring, stop repeating themes, cut length. Those are hard to encode as unit tests but easy to encode as recurring critique agents with veto power.
Teams can implement a lightweight pattern: generator agent plus critic agent with explicit rubric sections for delight, brevity, and novelty. Mollick played critic himself because built-in judges missed meta-quality.
How this interacts with Anthropic product strategy
Fable 5.2 sits beside Opus 5.5 in Mollick's praise. Routing cheap critique to one model and long-horizon execution to another matches public comparison posts on explainx.ai.
Anthropic's IPO roadshow does not change API behavior, but it increases scrutiny on AGI narrative. Mollick's restrained "almost" label is a useful external calibration.
Open artifacts you can inspect
Mollick published open repositories for Portals and the Auden recreation in his post replies, including GitHub links from his account. Those repos are useful ground truth when marketing clips outpace what you can reproduce.
Rebuilding a slice of Portals locally is a better eval than reading launch tweets. Likewise, stepping through Fate and the Individual pages tests literary UI craft agents claim.
Evaluation rubric you can copy
| Dimension | Pass signal | Fail signal Mollick saw |
|---|---|---|
| Factuality | Cross-model spot checks clean | Not primary failure mode now |
| Delight | User wants to continue experience | Joyless course first draft |
| Concision | Fits journal page limits | 182-page paper |
| Diversity | Varied themes across outputs | Repeated cyclone worlds |
| Self-stop | Agent proposes cuts | Rulebook bloat |
| External risk | No unsent customer mail | Prebid outreach paused by human |
Use the table in design reviews when vendors show cherry-picked demos.
Related on explainx.ai
- Top Fable 5.2 demos roundup
- Fable 5.2 launch pricing
- Frontier models compared including Fable 5.2
- Fable 5.2 code ownership prompting
- Google jagged frontier in science
- METR RSI covert evaluation notes
What to watch in competing frontier launches
Mollick predicts OpenAI and Google will ship answering models soon. When they do, rerun his rubric on taste failures, not leaderboards. Jagged AGI is a comparative state: if rivals duplicate Fable breadth but fix self-critique, the label expires quickly.
Until then, budget human hours for steering long jobs. Early access without metering, which Mollick discloses, flatters uninterrupted agent marathons. Paying customers will hit limits; design workflows for pause, resume, and human veto.
Subscribe to his One Useful Thing feed for methodology; subscribe to explainx.ai launch posts for vendor terms. Neither alone is enough for production choices.
If you run agent competitions internally, score presentations on Mollick rubric rows, not only task completion time. Completion without taste is how jagged systems ship to customers who cannot articulate why the product feels off.
Classroom and enterprise training angle
Mollick teaches entrepreneurship and has run public AI assignments for years. His Fable tests read like syllabus items scaled to frontier capability. Educators can reuse the structure: pick cross-domain prompts, require reflection essays on agent mistakes, and grade human fixes rather than raw model output.
Enterprise L and D teams can mirror that pattern in sandbox tenants. A week-long capstone where employees steer Fable-class agents on internal fake datasets teaches judgment faster than slide decks about prompt engineering.
Legal and compliance should review whether capstones risk exfiltration before copying university prompts verbatim. The pedagogy still transfers with redacted data.
Investor relations teams should treat Mollick "almost AGI" label as qualitative due diligence, not a revenue forecast. Capability demos impress; self-critique gaps predict services margin you will spend on human editors and reviewers. Budget those editors before you budget more tokens. Jagged AGI is an operating expense story, not only a benchmark story. Plan for that line item in 2026 roadmaps.
Mollick essay details are accurate as of November 10, 2026.