👋 Howdy to the 1,620 new legends who joined since our last edition! You are now part of a 572,972-strong tribe outperforming the competition together.

LATEST POSTS 📚

If you’re new, not yet a subscriber, or just plain missed it, here are some of our recent editions.

🍀 Happenstance: 10x Your Luck. An interview with Alex Teichman, Founder & CEO at Happenstance.
🚧 Selling Safety, Pricing Blind & Scaling Infrastructure AI. An interview with Shelley Copsey, Founder & CEO at FYLD.
🧑🏻‍🏫 Why Every CEO Needs Someone To Report To. An interview with Alexis d'Amecourt, CEO at Mochary Coaching.

PARTNERS 💫

Business banking was built for people to manage. Airwallex's agentic business accounts manage themselves. Powered by global

financial infrastructure, your AI agents can now provision local currency accounts, execute cross-border payouts, and optimize liquidity across entities 24/7. Put idle cash to work and move funds at the speed of your growth, all within the rules and approvals you define.

Meetings aren’t the problem. The challenge is everything that comes after them. With Granola, the AI Notepad for people with back-to-back meetings, it's no longer about scrambling to keep up.

Take notes the way you always have. Granola works in the background, turning conversations into clear summaries, action items, and next steps. Before, during, or after a meeting, you can chat with your notes to quickly understand what needs to happen, write follow-ups, or share context with others.

_{Interested in sponsoring these emails? See our partnership options here.}

HOUSEKEEPING 📨

Two of my favorite NBA players have been trending in different directions in my heart over the last month. First, LeBron James launched a partnership with the cultural cesspit, Polymarket. Shortly after, Victor Wembanyama, when asked if he’d ever do the same, gave his response: “Absolutely not. I will never do that. I think it’s very sad.”

I’ve been chipping away at a piece titled ‘How To Choose Your Heroes,’ where I plan to debut my Net Hero Score, a concept that functions as an NPS graded on a set of your values. This is a perfect microcosm of that idea. LeBron moved down my hero totem pole, while Wemby shot up. Seriously, how much money do you need to show some character? It broke my heart seeing LeBron selling his soul to a shitty gambling app.

BUSINESS STORY 🗞️

Zapier’s AI Offer: Valuemaxxing

22-year-old wannabe gigachad Braden Peters, aka Clavicular, recently popularized the term ‘looksmaxxing’ thanks to his incessant desire to ‘mog’ other men. To achieve this, he has undergone Gattaca-esque femur extensions, trimmed his nasal dorsum, and smashed his cheeks with a hammer—a practice unsurprisingly known as bonesmashing—in dreams of K-Pop-style cheekbones.

Somewhere along the way, the '-maxxing' suffix escaped the forums and made it to Silicon Valley. Zapier, the 15-year-old technology stalwart, coined (or at the very least, adopted) the term 'valuemaxxing,' referring to a company's ability to extract maximum value from our new intelligence overlords, Anthropic and OpenAI, largely by asking them to do less.

Zapier is in an enviable position in the AI harness wars because it's been doing this all along. Half the workflows we burn tokens on today don’t need to be agentic; they just need to be consistent. Zapier has always been consistent, and now it can layer on AI where necessary. To understand Zapier’s AI play, I sat down with Wade Burrell to discuss their product direction and vision for AI builders everywhere. Let’s jump into what I found.

Refresher: The ‘zap’ company

This might be the most redundant refresher I’ve ever given, but for anyone who’s been living under a rock for the last decade, Zapier (/ˈzæpiər/ ZAP-ee-ər) is a no-code and low-code workflow automation platform that connects more than 9,000 web applications. Founded by Wade Foster, Bryan Helmig, and Mike Knoop, and hailing from Columbia, Missouri, they are remote-first, have raised only ~$1.4 million, and are quite likely the only thing holding the majority of your company together right now.

If you’d like an elevator pitch for everything else of note Zapier has achieved in the lifetime of its sticky-taping-together business operations, I have compiled a helpful table for you below.

Not bad, right? Impressively, Sequoia and Steadfast bought secondary shares in Zapier at a $5B valuation in January 2021. That's roughly a 3,800x valuation-to-capital-raised ratio, the kind of number that makes VCs wet their pants with exultation. Zapier’s product, the ‘zaps,’ is great, but the OAuth handshake, token refresh, rate limits, and API management make it so sticky. But that’s what got Zapier here; it’s not necessarily what will get them there. For their next act, they'll double down by sitting on the fence.

The Harness Wars

If the battle of 2025 was ‘whose model is smartest,’ then the battle of 2026 is surely ‘whose harness do you live in.’ Harnesses being, of course, Claude Code, Cowork, Codex, Cursor, Gemini CLI, Notion AI, Perplexity Computer, Hyperagent by Airtable. Every company wants to own the experience you enjoy while building with AI.

While the battle for the harness rages on, today’s protagonist, Zapier, is doing its best to impersonate the country of Switzerland. Zapier MCP works with Claude, ChatGPT, Cursor, and all the aforementioned combatants. Your auth lives in Zapier, so the harness is disposable. As Wade (Wade Burrell, not Wade Foster; Zapier apparently has a Wade quota) would put it during our chat: “Our users are already in Claude, they're in Cursor, they're in ChatGPT. They're describing what they want in plain English, or writing code using Claude Code. MCP lets us connect to them, and lets anyone else connect to them as well. It gives us a way to meet them where they're already working.”

Here's the bit that makes the harness wars feel less like a tech story and more like a hostage situation: skills. Every skill your team has written lives inside one company's harness, tuned to one company's quirks. "If you have a business that was on Claude and you have all these skills you created for your entire org," Wade told me, "having to take all those and put them onto ChatGPT is a hassle."

Hassle is PMM for ‘a major pain in the ass.’ Wouldn’t it be amazing to have a central repository for all your skills instead of porting them from one model to another? Yes, yes, it would.

When tokenmaxxing died

Tokenmaxxing, another 2026 phenomenon, was the act of maximizing token usage as a sign of productivity. Meta ran an internal ‘Claudeonomics’ dashboard, Harvey went from one million tokens to 12 million a month, and even Jensen was claiming a $500k engineer should be burning $250k in tokens. This was all until the adults in the room, the CFOs, pulled the rug out from under them.

As quickly as you can say ‘the CFO would like to speak with you,’ the industry buckled over, regurgitated its ill-consumed tokens, and cried itself to sleep in the fetal position. What this looked like in practice: 1) Uber exhausted its 2026 AI budget in four months; 2) Salesforce racked up an expected $300M in Anthropic token spend across the same time frame; and 3) my head of finance got angry at me because I spent all of Athyna’s tokens in June having Claude rewrite the perfect ending to Game of Thrones.

It’s no surprise, then, that a recent survey of 396 organizations found that 40% escalated surprise AI costs to the board, 25% delayed or canceled an initiative, and 89% missed AI cost forecasts by more than 10%. A similar KPMG survey found that only 26% of companies fully understand their AI spend. Bonkers stuff.

Zapier wins by a mile here. Why? Well, put simply, Zapier uses AI only where it's required. Allow me to hand over to my trusted ally Wade to explain…

Wade: ”A tool call through Zapier costs tasks, which are a fraction of a penny. The same action through Claude can potentially burn tokens like it's nothing, and it doesn't necessarily need to. It just needs to do the thing that's in the step it has to do. When you talk about having sixty people in your company, and sixty people are doing that daily, doing tasks with AI that may not need to be done with AI, the bill just explodes. And that gap compounds fast."

Doc (me): “You have my attention. Please, go on, my bearded comrade…”

Wade: "Well, writing a record, or checking for a duplicate, or formatting a date, those things we have figured out and can be done programmatically. It's not something that benefits from creativity or from having to create something out of nothing, as AI often does. It just does the thing."

Wade (again): "You use whatever model you want where judgment actually helps you. Reading something unstructured, pulling out what matters, summarizing the whole thing, that's what AI does really well right now."

In layman’s terms, we are all wasting money hand over fist by using AI to redo all the steps we already automated with Zapier years ago. That brings me to my next learning.

Determinism beats probabilism

In this next section, I want to contrast deterministic versus probabilistic workflows. But before doing that, let’s step back and talk about MCP and what it actually does. Anthropic launched MCP, or Model Context Protocol, in November 2024 as an open standard for connecting models to tools and data. OpenAI adopted it in March 2025. Google followed suit in April of the same year.

Think of MCP as USB-C for agents. Something that you plug in and everything connects and works like magic. The launch of USB-C for agents led to the agentic explosion. All of a sudden, your Claude or ChatGPT could go and do things in the real world. Of course, a Cambrian-like explosion occurred. Take Zapier, with 450,000+ MCP servers and 18.5M tool calls. It’s the obvious extension from a helpful thinking partner to an agentic colleague.

So that’s all well and good. But what about the determinism-versus-probabilism conversation? Deterministic steps are those you’d be familiar with: a trigger followed by predefined actions. When the trigger occurs, the workflow runs its action steps. This is classic Zapier.

Claude, on the other hand, runs probabilistically: the model can decide which tools to use, in what sequence, and whether it needs to take another step, creating a loop of plan → act → observe → adjust. But almost every step of a typical ‘agentic’ workflow doesn’t need to be agentic at all; or at the very least, 80% of most tasks can run deterministically. Claude is awesome at probabilism; don’t get it twisted. Some workflows might split 50:50 deterministic: probabilistic, others 20:80. But the point is: why endure the cost, and (more on this next) the risk, if you don’t have to?

Risk, dear builder, is real. Risk of failure, or worse, a catastrophic mistake (more on this). Let’s look at some Year 11 maths (the last time I studied maths in school) to illustrate the idea.

Say a workflow has 20 steps, and each one has to succeed for the whole thing to work: pull the candidates, grab the feedback, check for duplicates, format the dates, write the summary, post to Slack, and so on. If a model gets each step right 95% of the time, that sounds fine. But the workflow only succeeds if step one AND step two AND step three... all go right. Right? So you multiply: 0.95 × 0.95 × 0.95, twenty times over. That's 0.95 to the power of 20, which comes out to 0.36. Your ‘95% reliable’ agent finishes the whole job about a third of the time. I assume this sounds familiar if you, like me, have your scheduled tasks in Claude break every other day.

A rule (an if/then, a database write, a date formatter) doesn't have a 95% or a 99%. It’s the same input, with the same output, every single time. So 1.0^20 is still 1.0. Chain twenty deterministic steps together and the workflow works every single run.

That’s not a fair comparison because it doesn’t use any AI (we want AI). So let’s look at 20 steps, with three running probabilistically (calling Claude). The 17 deterministic steps don’t contribute to the failure rate. Only the three model steps matter. At 95% per model step: 0.95 × 0.95 × 0.95 = 0.857. The workflow finishes about 86% of the time. At 99% per model step, it runs successfully 97% of the time.

These examples are illustrative, and of course, Claude might be more successful than 95% on average, but you get my drift. That’s the risk-of-failure side. Catastrophic risk is another thing entirely. PocketOS, for example, had an agent on Opus 4.6 delete the car-rental software startup’s entire production database and backups in nine seconds, leaving them 30+ hours down.

As you can imagine, the founder was one very unhappy camper (if only he’d read this email in the weeks leading up to the incident).

"You use AI for that judgment layer," says Wade, "which is interpretation, classification, drafting. Then the deterministic automation for the repeatable execution, so triggers, routing, data writes, formatting, approvals, and system updates."

Put simply, determinism buys you fewer failures, lower bills, and a workflow that behaves the same for every person on the team. Cost is the headline, but consistency is the real product.

The single model handcuffs

I called Anthropic and OpenAI our intelligence overlords earlier. Fair. But if you've watched the throne this year, the succession has been a bloodbath. Anthropic alone released nine frontier models in ten months. OpenAI shipped three GPT-5 point releases in the same window. Zapier's own AutomationBench leaderboard reshuffles so often that whoever's on top as I write this will probably have been knocked off by the time you read it. Even Wade's own pick for marketing work is Gemini Flash first and Opus second, and he works at the company that runs the leaderboard.

This is to say that when you are stuck inside one model’s harness, you have sworn fealty to whoever is on the throne this week. Those who use an interoperable harness will produce higher-quality output consistently.

And while I’ve waffled on about pricing in earlier sections on killing off the steps that don’t need to be AI, a multi-model harness also offers colossal savings. Anthropic’s Fable 5.1 costs $10/$50 per million tokens (the first price is for input, the second for output), while Opus 5 only costs $5/$25. Gemini 3.8 Flash is lower again at $0.75/$3.75. That means the same task could cost up to 13x less, depending on which brain you point at it. Not inconsequential numbers. Do you want to compete with that? I certainly know my answer.

But you don’t need to worry only about cash flow. Anthropic cut Windsurf's Claude access entirely with under five days' notice in June of 2025. That August, it revoked OpenAI's API access outright. In April this year, Anthropic's automated systems suspended the accounts of Argentine fintech Belo (Argentina being my old stomping ground), locking about 60 staff out for 15 hours before calling it a “false positive." Just last month, OpenAI cut Cursor off after the SpaceX takeover. None of those companies did anything wrong; they just built on one brain. And one brain, it turns out, can be switched off by a policy team on a Tuesday.

How I yearn for multiplayer

My final gripe with most harnesses today is that they aren’t multiplayer. And it’s hurting the industry. An MIT study from last August found that 95% of custom enterprise pilots fail, while 90% of employees use personal AI tools for work, and only 40% of firms had bought a company subscription.

What this means in reality is if Jill goes away on holiday, and 80% of the creative team’s important workflows are built inside her Claude account, her colleagues need to: a) rush to rebuild everything inside their own accounts, or b) even worse, regress to the Stone Age handling of all the automated tasks manually. Uggggh.

An even more painful thought experiment is what happens if Jill leaves and her offboarding isn’t sufficiently awesome. All those magic AI tricks walk out the door with her. Zapier MCP, on the other hand, makes all your connections, permissions, and workflows team assets managed from one place. Which, funnily enough, is what SaaS was always for. AI infrastructure is turning back into enterprise software, and that's a good thing. Good, old, deterministic, multiplayer software.

Playbook / how you can apply this

If you were so inclined after this reading to want to get off the single model superhighway, then these are the easy steps I’d follow.

• Install Zapier into the AI you already use: Connect Zapier MCP to Claude or ChatGPT, scope permissions per app, and never hand an agent a password again. Win.

• Audit for judgment vs execution: Use models for reading unstructured stuff, classifying fuzzy things, drafting. Stay on rails for triggers, routing, writes, formatting, approvals, system updates. If a step runs the same way every time, it never touches a model.

• Keep a human on anything irreversible: Drafts, not sends; approvals before deletes and so on. Ask the PocketOS founder how the alternative worked out for him.

• Rent the brain, own the hands: Keep auth in Zapier, be ready to swap out your models when the leaderboard flips, and rest easy knowing nothing is going to break when your colleague takes a holiday.

Future

This is not the first time I’ve spoken deeply about harnesses in this newsletter, and I don’t expect it to be the last. The more I think about AI and how we operate it, the more certain I am that Anthropic or OpenAI won't win the harness wars. Living inside one lab's harness just isn't a good enough experience. Lots of downside, very little comparative upside.

We are still so early it's impossible to know who will come out on top, but Zapier is making a very compelling argument. What I can tell you after writing this piece is that for the foreseeable future, if you need me, you'll find me over in the corner, valuemaxxing like my life depends on it.

Fun facts

• The side-hustle years: All three founders kept their day jobs for the first year. Wade Foster was doing email marketing at an insurance software company, and Bryan Helmig was freelancing while they built Zapier nights and weekends

• The $100 PayPal: Andrew Warner paid for Zapier before it launched. Wade Foster built a Zap to text himself when it landed. It buzzed as they walked into the bank to open the business account.

• More agents than employees: Wade Foster confirmed on the Madrona podcast in April that Zapier now runs more AI agents than humans on payroll. Roughly 800 people, and counting the other lot is harder.

• Say it right: I’m told by a very reliable source that it's ZAP-ee-er, as in ‘Zapier makes you happier.’ Fifteen years in and I wasn’t sure myself.

Extra reading

• Building Athyna Intelligence - January, 2026

• The World’s Hottest Industry: Post-Training - March, 2026

• Zapier Has More AI Agents Than Employees - April, 2026

And that's it! You can follow me on Twitter and LinkedIn, Wade on LinkedIn, and also don’t forget to check out Zapier MCP while you’re at it.

BRAIN FOOD 🧠 

TWEETS OF THE WEEK 🐣 

TOOLS WE RECOMMEND 🛠️

Every week, we highlight tools we like and those we actually use inside our business and give them an honest review. Today, we are highlighting Framer*—the site builder trusted by startups to Fortune 500.

See the full set of tools we use inside Athyna & Open Source CEO here.

HOW I CAN HELP 🥳

P.S. Want to work together?

1. Hiring global talent: If you’re hiring tech, business, or ops talent and want to do it 80% less, check out my startup, Athyna. 🌏

2. See my tech stack: Find our suite of tools & resources for both this newsletter and Athyna here. 🧰 

3. Reach an audience of tech leaders: Advertise with us if you want to get in front of founders, investors, and leaders in tech. 👀