Preview: All About Agents — Dots Arrive Before OpenAI Is Ready
Summary
- OpenAI’s Dots launch exposes a positioning mismatch: it looks like a friendly mass-market product but behaves like a narrowly useful, unfinished productivity agent. Andrew calls the result “confused and confusing”; Ben sees the missing coordination layer Codex needs, yet finds Dots “weirdly neutered” and unaware of his actual projects.
- Ben had largely written off OpenAI’s consumer prospects after Gemini 3, but Amazon may give the company another route to monetization. OpenAI could become an agentic surface for Amazon products and ads: “Access our products, no problem. They also show our ads. That’s the trade.”
- Consumer agents should be sold as pain relief, not as accessories for the “grind lifestyle.” Ben’s pitch is the “modern-day Tylenol, the modern-day Advil”: technology has become painful for normal people, from roughly 700 apps to logins, two-factor authentication, and device migration.
- Effective agents need fresh, task-specific context and durable external records—not one enormous conversation. Ben argues that million-token context “sucks” when old instructions become buried; his preferred system continually writes things down, retrieves the relevant material, and starts a new thread when a task changes direction.
- Claude and Codex currently occupy complementary roles in Ben’s workflow. Claude is the context-preserving “chief of staff,” while Codex is better for executing a defined project—Andrew’s analogy is that “Claude is the barrel, and Codex is the ammo.”
- OpenAI’s larger mistake was forcing ChatGPT, ChatGPT Work, and Codex into a super app before their architectures converged. Ben thinks the destination is right but “they had no patience”; agents will ultimately beat apps because users can state what they want without knowing how to do it, and “it’s going to win because it’s going to be better.”
Deep dive
1. Dots arrived amid a welcome return to the AI Wild West
Andrew’s emblem for the moment: after OpenAI announced Dots, SpaceX AI bought dot.com and redirected it to the Grok download page. Ben welcomed the prank as a revival of the era when tech companies publicly mocked competitors—behavior harder to sustain once the industry became “a bunch of monopolies who have their own little fiefdoms.”
Listener Steven supplied the sharper launch critique: OpenAI has 90% consumer mindshare, yet showcased “San Francisco-coded techno-parents” building startups and side hustles instead of relieving family scheduling, kids’ activities, catching up with friends, groceries, or caring for parents. Andrew found the branding similarly incoherent: cuddly characters and a lowercase “d” signal a mass-market consumer product, while the demonstrated product remains productivity-heavy.
2. Consumer agents need to remove pain, not glorify work
Ben admitted he had “given up on OpenAI in the consumer market ages ago,” particularly after Gemini 3 looked strong and OpenAI still lacked an advertising model. He softened that conclusion because Amazon benefits from OpenAI as a surface for its external agentic opportunity: OpenAI can access Amazon’s products while also showing Amazon’s ads.
Andrew’s household offered a small counterpoint to OpenAI’s positioning. His wife initially asked why she needed Muse when ChatGPT already handled her personal organization, but after trying it for two days, “she’s now a Muse user.”
Ben’s preferred framing is unequivocal: “market these as the modern-day Tylenol, the modern-day Advil.” Agents matter because ordinary technology is painful, not because agents are novel; apps improved on what came before, but that never made them the final interface.
The concrete pain is cumulative: roughly 700 apps, forgotten locations, recurring logins, two-factor authentication, and device migrations. Ben abandoned moving to a Max halfway through because reversing the migration would mean enduring the same ordeal again: “It’s awful.”
3. Agent quality depends on context architecture
Codex’s project-and-thread structure maps cleanly to folders on Ben’s computer, making it effective when he knows exactly what he is building. Its willingness to discard a failed direction and start a fresh thread is a feature: carrying every mistake forward only confounds later reasoning.
Ben’s reversal on context windows is notable. The early assumption was that “larger and larger context” would solve everything; his current view is, “Actually, no, that sucks.” Long threads bury initial instructions, while capable agents continually write state down, keep active context fresh, and reload only what becomes relevant.
Claude handles continuity better, so Ben uses it as a “chief of staff” spanning servers, cameras, and other projects, then shifts defined execution into Codex. He has his agents write to ordinary files outside
.codexor.claudebecause he is “allergic to locking.” Andrew’s summary landed: “Claude is the barrel, and Codex is the ammo.”
4. Dots identifies a real coordination gap but does not fill it
Conceptually, Dots looked like the overarching liaison Ben wanted for his Codex agents. In practice, “Dots kinda sucks”: it ignored his projects, discovered a calendar plugin, and fixated on an October 14 conflict—prompting both “How did you know that?” and “That’s not what I need help with.”
Dots also failed to recognize Ben’s custom Codex permissions file or where it was until explicitly directed to it. That made it simultaneously intrusive about irrelevant information and “unknowledgeable” about the work it was supposed to coordinate.
Spaces, by contrast, resembles a system Ben independently built for Gecko: cards hold an item’s history and attachments, including an interview transcript ready for editing. Even the interface can remain lightweight—clicking Done merely sends Gecko a prompt, and the agent performs the bookkeeping.
The frustration is that many constituent ideas “make sense and are super cool,” but no organizing principle makes Dots, Spaces, Pages, scheduled tasks, ChatGPT Work, and local versus cloud execution legible. Compute constraints deepen the mismatch: Dots could help broader audiences, yet initially reaches only Pro-tier subscribers.
5. OpenAI’s super-app vision arrived before its foundations
Ben calls the “cardinal sin” the super app. ChatGPT was stateless across fresh conversations, ChatGPT Work ran inside a VM, and Codex operated on the local computer; combining them while they still had different architectures “didn’t make sense architecturally” or conceptually. “They had the right vision, but they had no patience.”
The intended destination still convinces him: users should state what they need rather than learn which app, website, or workflow performs it. Ben says agents will win because the experience will be better—not because nerds favor them or the economy demands them. Andrew frames that advantage as a consequence of the rest of tech sucking. The preview ends just as a listener challenges whether consumer agents can be winner-take-all when demand does not improve the product and switching costs remain structurally low.