Ethan Mollick Let Codex Loose on His X Bookmarks for Two Hours — Here's What It Reveals About Agents

How long can a coding agent run unsupervised, and what can it actually finish in that time? On August 16, 2026, Wharton professor Ethan Mollick (@emollick) ran an experiment that answers both questions with real numbers instead of a benchmark score.

He asked Codex — OpenAI's coding agent, running with computer-use browser automation directly on his own machine, logged into his own Chrome — to download all of his X bookmarks and organize them into "an easy-to-sort file, like Excel," with text, links, categories, and whatever else seemed useful. He told it to feel free to spin off sub-agents. There is no official bookmark export button on X, so this meant scraping a page, not calling an API.

Codex worked for 1 hour, 46 minutes, and 54 seconds, unsupervised. It came back with: "Export complete: 5,032 unique X bookmarks in a filterable Excel workbook."

This is useful as more than a party trick. It's a live data point on what long-horizon agent loops and browser-driving coding agents like Codex's computer-use mode or Claude in Chrome can now do end to end — and a live data point on the risks people flagged the moment it went semi-viral.

TL;DR — What people are asking

What Codex actually did in 1h 46m 54s

The prompt was simple in scope but open-ended in execution: export every X bookmark, structure it for sorting in Excel, and use sub-agents if useful. No step-by-step instructions on how to scrape a page nobody officially lets you export.

The finished workbook, saved as X_Bookmarks_Export_2026-08-16.xlsx, contained:

• Post text and timeline excerpts — authors, dates, engagement metrics, and URLs for all 5,032 unique bookmarks
• Classification layers — primary categories, topic tags, AI subtopics, and short explanations of why each item was classified that way
• 8,372 normalized link and media records — every URL and media asset inside every bookmarked post, deduplicated and structured
• Media importance labels and available alt text
• Five sheets — Summary, Bookmarks, Links & Media, Categories, and a Read Me explaining the schema

Codex also documented X's own data limitations inside the workbook rather than glossing over them: some long posts only show up as collapsed excerpts, most external links remain unresolved t.co redirects, video URLs are thumbnail-only, and X exposes neither bookmark timestamps nor a definitive total bookmark count anywhere in the UI — so Codex couldn't confirm 5,032 was literally everything, only everything it could reach.

That last detail matters more than it looks. An agent that reports "5,032 bookmarks, done" without caveats is making a claim it can't back up. An agent that reports "5,032 unique bookmarks, and here is exactly what X's own interface won't tell either of us" is doing something closer to real data engineering — self-verification and honest uncertainty, not just task completion.

The second task: 3 minutes, and a completely different kind of work

Once the export was done, Mollick asked for something qualitatively different: "find some extraordinary gems among these."

That's not an extraction task. There's no ground truth to verify against — "extraordinary" is a judgment call. Codex took 3 minutes and 11 seconds and came back with a curated shortlist, explicitly selected "for surprise, durability, and rabbit-hole potential — not raw popularity," split into three groups.

Intellectual surprises:

• A bookmark, sitting at only 15 public likes, about sperm whales apparently sharing information about whaling attacks — harpoon success rates reportedly fell 58% within a few years, framed as a case of nonhuman cultural adaptation
• A 500,000–300,000-year-old handaxe built around a fossil shell — potential evidence of extremely early aesthetic intention
• A Michael Levin paper/thread on the underestimated collective intelligence of biological systems
• A thread on why Wikipedia succeeded where seven similar knowledge-collaboration projects failed
• A bookmark on "fooled by beautiful data" — attractive visualizations earning more trust than they've proven
• A thread on "how to pick a research problem"

Wonderful human rabbit holes:

• A man who spent five years recreating all 107 recipes in the Alinea cookbook
• The Financial Times turning the opening of WWII into a map-prediction game with cash prizes
• Darwin's reply to a critic of On the Origin of Species — reportedly Julia Galef's favorite letter in the history of science
• The technical debt hidden inside Dungeons & Dragons' time and distance rules, inherited from Chainmail
• Avant-garde perfume as a cheap experimental art form

Things worth opening and playing with:

• ShadeMap — simulates every mountain, building, and tree shadow on Earth for any date and time
• The Whole Earth Index — a near-complete free archive of Whole Earth publications
• "54 African countries, 54 architecture projects" — architecture as cultural history
• Infinite Craft, Neal Agarwal's endless combinatorial crafting game

Codex's own top three: the whales, the fossil-shell handaxe, and the Alinea project — each, in its words, containing "an entire essay hiding inside one bookmark."

Why the two-task split matters more than the total time

Read the two numbers side by side: 1h 47m to extract and structure, 3m 11s to curate. That gap isn't an accident of task size — it reflects two genuinely different kinds of agent work.

This is the same distinction that shows up across agent harness engineering generally: extraction and transformation tasks scale with sub-agent orchestration and more wall-clock time, because you can throw parallel workers and verification loops at them. Judgment tasks don't scale the same way — they compress, because the model already "knows" what's interesting once it has the structured data in front of it. Mixing the two into one prompt would have made both worse: a scraping job padded with taste, or a curation pass buried under classification busywork.

If you're designing your own long-horizon agent loops, this is the practical takeaway to steal: split extraction from judgment as separate prompts or separate stages, even in the same session, rather than asking for both at once.

The part worth being honest about: this probably wasn't clean

The replies to Mollick's thread raised real, specific concerns — and they deserve equal billing with the "wow, agents can do this now" framing, not a footnote.

Terms of Service risk. Brian (@BrianG12321) noted he was surprised X didn't block this — browser automation driving a logged-in account this way is explicitly against X's Terms of Service. He also pointed out that X's official API pricing is prohibitively expensive for something as small as personal bookmark access, which is part of why people reach for workarounds like this one in the first place.

Bot-detection / account risk. efe (@extliqprovider) warned that letting an agent drive a logged-in browser session this way risks the account getting flagged as bot activity — a real, asymmetric downside: the agent doesn't bear the cost of a flagged or suspended account, the human does.

A safer path already existed. Oliver Schaudt (@oshout) pointed out that X's own data-export tool — the "Download an archive of your data" feature, similar in spirit to a GDPR data request — already includes bookmarks and likes. That raises a fair question: was nearly two hours of unsupervised agent time and browser automation risk actually necessary here, or would the built-in export plus some spreadsheet cleanup have gotten most of the way there with none of the ToS exposure? X also runs a hosted MCP server with a documented bookmarks endpoint for apps with proper API access — a sanctioned integration path that existed the whole time, separate from both the consumer export and the scraping route Codex took.

Usage-pattern reality. Kevin Wimer (@kevinace) made the driest and maybe most honest point in the thread: 99% of things bookmarked on X never get looked at again. That's not a knock on the engineering — it's a fair question about return on the roughly two hours of combined agent-plus-review time this took, versus just... not doing it.

Model and accuracy questions. Harrison Bergeron (@BigLawNoMaw) asked what model and reasoning level were used and how accurate the output actually was. Mollick's own reply thread is informative here: he'd tried Grok's bot first, but it lacked API access to bookmarks at all, and separately noted the plain X API "doesn't do all the bookmarks apparently" for some reason. Seth Saler suggested Grok's X plugin might have read access that could export bookmarks too, but Mollick said Grokbot does not have bookmark access via the API. Chris Dilger (@cdilga) mentioned other people getting similar bookmark data through the official export tool combined with third-party indexing tools — reinforcing that Codex's browser-automation route, while it worked, wasn't the only route.

None of this erases what the run demonstrated. But a fair accounting of "what long-horizon agents can now do" has to include "and here's what it cost in platform risk to find out."

What this is actually a case study in

Strip away the specific bookmarks and this is a compact demonstration of four things agent builders care about right now:

1. Long time horizons without babysitting. Almost two hours of continuous, unsupervised operation on an open-ended task, with no checkpoint back to the human until it was done. That's a meaningfully longer leash than "watch every tool call."
2. Self-verification as a built-in step, not an afterthought. Codex reported checking its own formulas and visually reviewing every sheet — and documented what it couldn't verify (X's undisclosed bookmark timestamps and totals) instead of silently pretending completeness.
3. Sub-agent orchestration on request. Mollick explicitly permitted spinning off sub-agents; the workload — scraping, classifying, deduplicating 8,372 link records — is exactly the kind of parallelizable work that benefits from it, in the same spirit as multi-agent orchestration patterns showing up across other coding-agent harnesses.
4. A distinct judgment layer after the mechanical one. Curation isn't a bigger version of extraction — it's a different skill, and Codex handled the switch in the same session without being told explicitly how to switch modes.

That combination — hours-long autonomy, self-checking output, orchestration, and a follow-up judgment task — is a reasonable snapshot of what coding-agent harnesses like Codex, Claude Code, and their peers are being asked to do in production-adjacent settings by mid-2026, well past "autocomplete for functions."

Practical takeaways if you're designing similar workflows

If you build with agents — Claude Code, Codex, or any harness with browser or computer-use capability — a few things from this run are worth carrying into your own designs:

• Check for an official export or API path before automating a logged-in browser session. X had one (the data-export archive) and a sanctioned integration route (hosted MCP with a bookmarks endpoint). Reach for scraping only when the official path is genuinely missing the data you need, the way Mollick found the plain API doesn't — and know you're trading convenience for platform risk when you do.
• Understand ToS and bot-detection exposure before you start, not after. Browser automation against a logged-in personal account can get that account flagged — the agent doesn't pay that cost, you do.
• Separate extraction from judgment as distinct stages. Ask for the structured, verifiable output first; ask for taste, curation, or "find the good stuff" as a second, separate prompt once the agent has clean data to reason over — as this run's 1h47m/3m11s split shows.
• Ask agents to document what they can't verify, not just what they found. "X doesn't expose bookmark timestamps or a total count" is more useful than a confident number with no caveats.
• Budget the human review time, not just the agent's runtime. Nearly two hours of unsupervised agent work still needs a human to read the output — factor that into whether a workflow is actually worth running, especially against Kevin Wimer's blunt point that most bookmarks never get revisited anyway.

Related reading

• What Is Loop Engineering? Long-Horizon Agent Loops
• Claude in Chrome: Browser Extension Guide and Real Risks
• OpenAI Codex Computer Use on Windows: Mobile Steering, Parallel Worktrees
• Ethan Mollick: Prompting Tricks Are Over — Wharton Prompting Science
• X Launches Hosted MCP Servers: Bookmarks, Search, and More via API
• Agent Harness Engineering: Terminal-Bench and Orchestration Patterns
• Is Claude Cowork Safe? Security Vulnerabilities Explained
• What Are Agent Skills? Complete Guide
• Update — August 18, 2026: Mollick's father used Codex to set up his new MacBook with zero coding — see How to use Codex to set up a new computer

Details in this post reflect Ethan Mollick's public posts on X from August 16, 2026, and reply threads visible at time of writing. X's Terms of Service, API pricing, and data-export features may change — verify current terms before automating any logged-in session.