Most marketers use ChatGPT to do general research and then call it an AI strategy. The ones outperforming them are deploying specialized agents built for specific jobs.

We put together 22 plug-and-play ChatGPT marketing agents that handle the work eating your week, each with built-in instructions and structured outputs ready to go in under 5 minutes.

Subscribe to Marketing Against the Grain and get all 22 free.

Inside you'll find:

Your competitors are already running agents like these.

Get 22 ChatGPT Marketing Agents free when you subscribe to Marketing Against the Grain today.

Get The Guide

Ian Silber, head of design at OpenAI, has cross-workforce survey data showing designers and user researchers are the most overwhelmed, most anxious, least optimistic and most tired cohort in tech, and he names the cause himself: engineering got 100x faster while the design process still takes the time it takes. Klaviyo's CEO, Travis Kalanick, Susan Kare and a growth lead who scaled Wispr Flow converge on the consequence - the constraint has moved from hands to attention - and this brief works out what that means for staffing, spend and sequencing.

💡 Top tip: if you manage designers or researchers, read sections one and three today and leave the rest for the week. If you own a growth number, jump straight to sections four and five.

This week's brief synthesises 13 conversations into one theme; where two shows feature the same guest, we've merged them.

Featured in this round-up:

🕒 Estimated Reading Time: 10 mins. Time saved: 748 mins! 🔥

🚪 Not feeling this one? Five from the archive that might land better:

Three genuine disagreements, and one convergence that arrived from four unrelated directions.

The first is whether design work can actually be compressed - Silber says the process resists it; Bialecki mandated agent fluency across all 2,300 people at Klaviyo and reports "very little pushback on this." The second is whether there is a data ceiling on machine science - The Generalist says the real observations are trapped in unpublished raw data; Brian Greene sees no barrier in principle. The third is when to switch paid on - Stebbings states the orthodoxy, Swulinski says day one.

The big tensions to watch:

———————————————————————————

Lenny Rachitsky's cross-workforce sentiment survey has an unambiguous result, and he put it straight to Ian Silber: designers and user researchers score worst on every dimension measured - "the most overwhelmed, the most anxious, the least optimistic, the most worried, the most tired." Silber, who runs design at OpenAI, does not dispute it. He supplies the mechanism: engineers have 10x'd or 100x'd their productivity, but the design process still takes the time it takes - you try a hundred things and throw ninety-nine away.

Bialecki's answer at Klaviyo is a mandate, not permission. All 2,300 people, designers included, were expected to be running teams of agents by end of June - and he reports almost no pushback. Bialecki is describing what firms are doing to the role; Silber is describing what it feels like to be inside it while the definition changes under you. The gap between those two postures IS the survey result.

Key Takeaways:

Four conversations reached the same conclusion this week without touching each other: the scarce resource is no longer people who can produce output, it is people who can direct it. Kalanick puts it most compactly: "the only constraint to our imagination is management capacity. So then the question is twofold. Do you have an imagination and is it any good?"

Chai Discovery arrive at the identical point from biotech - with building now cheap, engineers become "allocators of attention". Swulinski gives it a timeline: within three years companies operate like boards of directors, 20% strategy, 80% agent execution. The implication is uncomfortable: you can hire your way out of a capacity constraint, but not out of an attention constraint, because adding people consumes the thing that is scarce.

Key Takeaways:

Susan Kare arrived at Apple with an art history PhD, no computing background, and bitmap sketches on graph paper that now sit in MoMA. Her YC talk is about the people who taught her judgement - chiefly Paul Rand, who told her to make things meaningful and memorable, and refused Steve Jobs' five-firm logo contest: he would take $100,000, produce one NeXT logo, and they would like it. They did.

Cadi Zhang describes the same boundary from inside a 2026 toolchain. AI-generated Blender assets came out approximate and wrong, so she draws the asset on paper first and only then hands it to the model - front-loading the human specification before generation starts. Silber turns the pattern into a market prediction: "when anybody can make anything, that's going to become one of the most important dimensions of a product." Generation collapsed the cost of approximate. It has not touched the cost of exact - and exact is where differentiation now sits.

Key Takeaways:

Matt Swulinski's read on the talent market: it has already bifurcated and most companies have not repriced. Tobin, who built Wispr Flow's viral UGC programme at 19 before moving to Polymarket, is his exhibit - "you would pay through the nose for talent like Tobin who just gets it." He predicts marketing acquisition gets its ML-researcher compensation moment within a year.

Chai Discovery describe the inverse - talent obscurity: not enough capable people enter biology because the work is illegible, so they are making the domain visual - a recruitment strategy dressed as design. Hard Fork supplies the market-clearing evidence: Korean chip engineers on $400-500k bonuses are now the most sought-after people in the country's dating market. The premium is not tracking credentials - Tobin is 19, Kare had no technical background. It is paying for judgement that survives contact with an undocumented system.

Key Takeaways:

Swulinski's central claim: SaaS marketing has been running the wrong template. "My philosophy is the e-com playbook is the right playbook for SaaS" - every cent maps to a conversion, fed by hundreds of UGC creators and 400-500 new creatives a month. The provocative part is timing. Stebbings puts the orthodoxy to him - paid is a dangerous early habit - and Swulinski's answer is: switch it on right away, because paid is the fastest instrument you own for validating positioning.

He is not arguing for dependence - healthy organic sits at 35-45% of acquisition. Where Silber and Kare would locate the moat in craft, Swulinski will not have it: "distribution to me is the only moat." Read closely, though, his actual claim is that paid is a testing instrument first and a channel second - a research argument, and a faster one than the discovery cycle Silber says no longer has time to run.

Key Takeaways:

Andrew Bialecki has started treating the agents Klaviyo builds as a customer segment: "We've actually started asking the agents we built what functionality is holding you back." The example is the most immediately usable thing in the week. Google's decade-old AMP spec for interactive email died because building for it meant constructing a web app inside an email payload - Klaviyo's customers could not use the raw HTML editor. "Almost on day one, our composer agent figured it out." A feature shelved for ten years on usability grounds became viable the moment the user stopped being human.

His infrastructure advice: agent-friendly APIs, documentation written for agents, and an outside agent representing the business to whatever is asking. Cadi Zhang raises the case he does not address: environments full of agents that are not aligned with your intent, including the detection counter-industry Hard Fork covered. Bialecki's optimism holds for agents you built. Zhang's caution applies to every agent you did not.

Key Takeaways:

The Generalist's episode contains the clearest statement of the limit on machine science anyone has offered recently: "There's going to be a limitation on how much you can learn about science by reading about other people doing science." More of the observations that matter still sit in raw instrument data than have ever been written up - frontier models are trained on the write-ups.

Chai Discovery confirm it from the bench: their dream fix is the validation loop - knowing instantly whether a designed protein works. Brian Greene disagrees on principle: asked whether a machine could produce something absent from its training corpus, he sees no barrier. But the two positions run on different clocks - Greene is describing decades; The Generalist and Chai are describing what the corpus contains this quarter. Only the second is actionable.

Key Takeaways:

Hard Fork ran the Airtable numbers: valued at $11.7bn in 2021, just acquired by Bending Spoons at a $1.29bn enterprise value - and an approach from Bending Spoons is itself a market verdict. Johann Hari's GLP-1 account is the cleanest available model of value migrating when a technology changes behaviour rather than products: Krispy Kreme downgraded, airlines briefed on thinner passengers, Nestlé nervous - nobody did anything wrong, customers simply stopped wanting as much.

Chai Discovery add the number that should unsettle anyone reading valuations as importance: until roughly three months ago, GLP-1 revenue exceeded all AI lab revenue combined, on a fraction of the market cap. Decoder asks whether Google even wants to win the frontier, or whether selling TPUs and cloud is simply the better business. The planning question: which curve is your roadmap priced against - the attention curve or the revenue curve? This week they are visibly diverging.

Key Takeaways:

Charity Majors has watched engineering organisations split into enthusiasts and sceptics, and her position is that the split is not ignorance on either side: "The problem is that neither side is making it up." The sceptics are grappling with real, worsening problems; the enthusiasts are correctly reading a race.

Her operating advice contains the line most teams have not budgeted for: "You got to get worse before you can get better." Not wandering is not an option either - you will be left behind. Her own organisation needed a year of the journey before she could trust a colleague saying AI is the wrong tool for a job.

Key Takeaways:

Kalanick endorses the maxim in the Founders episode title without hesitation - his illustration is a world-class marathoner at mile 21, emphatically not smiling. The disposition came from four unsalaried years before Uber. Then comes the revision, the most useful thing he says in either appearance: "I was running a $70 billion company the way somebody who thought he was going to starve next week would run it." He stands by the decisions but concedes the operating posture. Today he runs a few inches off that line.

The rest is the Adams thesis: everything in a city has been grown, mined or manufactured, each a multi-trillion dollar industry awaiting automation - automate a gold mine end-to-end and what is a mining company but real estate? Majors adds the external pressure: X runs on a fifth of the previous headcount - hold your own standard or someone imposes it. The revision matters because it separates the intensity that keeps a startup alive from the operating mode that survives at scale - and most leaders absorbing an agent mandate are being asked to run the first one indefinitely.

Key Takeaways:

———————————————————————————

Earlier Pod Shots that build on, contradict, or extend this week's theme:

Three unresolved questions from this week's brief worth carrying into your team:

That’s a wrap.

As always, the journey doesn't end here!

Please share and let us know what you liked or want changing! 🚀👋

Alastair 🍽️.