Disclaimer: This article does not discuss the five-part masterpiece by Pink Floyd, or your household pet. I’m still working on the words to describe the beauty we hear in those lead lines and the joy we feel when our best friend shows excitement due to our presence alone. I may never succeed in that task, because my perception is my own. Not yours, and certainly not anything that an LLM can truly understand.

The Six-Dollar Chair

In 2010 I was nineteen, a college dropout, and I got my haircuts at a school whenever a relative would say “boy, you’re looking shaggy today”. I didn’t (and still don’t) wear long hair well. EQ School of Hair Design in Council Bluffs (now SOHO Hair Academy ) charged six dollars if you let a student cut while an instructor watched. I could stomach six dollars, even as a part-time retail worker.

One afternoon the student finished, stepped back, and waved her instructor over, proud of the work. He leaned in and found a single hair she had missed along the neckline. He did not just snip it and send me on my way. He turned to her and said, “Honey, you missed one here. Men are like dogs. If you give them what they want, they’ll always come back. Reliable quality is the goal. If they go home and their girlfriend tells them something’s off, they’ll get their next haircut somewhere else.” She fixed it. He checked it again. I came back three weeks later for the exact same thing.

The instructor’s words have stuck with me ever since.

Back then I made $8.50 an hour at Menards , a 17.24 percent raise over the $7.25 I pulled at Pizza Hut in high school. (Iowa’s minimum wage is still $7.25 today, in 2026, but that is a different essay for a different day.) I drove a beat-up $900 Ford Contour I bought from a man with a sign that said “Pitbull with AIDS” in his front yard to deter would-be thieves. I kept my boxes of CDs, and concert ticket stubs in a canister wrapped in a Super Mario wristband I had purchased from Hot Topic (oh, what a time) - my proof of experience. At Menards I wore a blue vest and sold kitchens, building cabinet and appliance layouts in AutoCAD through an absurd run of clicks and keystrokes until I could turn the monitor around and show a homeowner a render of what their kitchen could become. I was already doing by hand, slowly, the thing AI now does in a second: turning a vague want into a finished picture. I just did not know yet that the picture was the easy part.

The Lesson

The instructor was not really talking about hair. He was teaching that quality is the discipline of finding the stray hair before the customer’s girlfriend does. The work is finished only when there is nothing left to catch. I didn’t have a girlfriend at the time, so I’m not certain anyone would’ve caught it for me, but I’m grateful for his words.

I heard the same lesson in three accents in the years that followed. Tom Groepper ran the busiest Menards in the chain, and he still does (at least as of my last visit to their newer, bigger location on the other side of town). He told everyone the same thing: “The customer is here to solve a problem. Your mission is to solve it in one shot. If they ask for drywall screws, send them home with the mud they need to patch that hole in their wall.” At a second job at Cabela’s, a manager put it another way: “These aren’t needs, they are desires. Understand what the customer truly wants before you recommend a product to them.”

Three jobs, one lesson. Think of the screws as information. The mud as judgment. Taste is the sense of sending the customer home with drywall mud to patch the hole left behind by the screw, so that he doesn’t have to make two trips on a Saturday to fix the wall in his home. None of it arrived as a download, or was purchased through a twenty-dollar a month subscription. It came from someone older standing over my work, catching what I missed, then letting me try again. Allan Collins and John Seely Brown named that loop cognitive apprenticeship in 1989: model, coach, scaffold, then fade. That instructor faded, but his words never did. Three weeks later I judged the back of my own neck in the mirror and caught the stray hair myself. The customers at these stores continue to come back, because they trust that their needs will be met, at a price that feels fair to them.

That is taste: repeatable judgment about quality you can run before anyone hands you the answer. A lot of sharp people are arguing right now that this faculty is the new moat, the scarce asset once making things gets nearly free. They are right, but they skip where taste comes from, and what happens to it when we automate away the chair, the instructor, and the stray hair.

Crutch or Coach?

There is always slop, and there is far more of it now. What bothers me about this wave is not the quality of the prose. Most of it is perfectly competent. It is also tiring, a haircut with the stray hair left in, printed millions of times per day.

The headlines say the tools are making us worse, and they are not all noise. In June 2025 an MIT Media Lab team led by Nataliya Kosmyna published “Your Brain on ChatGPT,” and the group writing essays with a language model showed the weakest neural connectivity of any condition. The authors called it “cognitive debt.” Microsoft and Carnegie Mellon University researchers surveyed 319 knowledge workers in 2025 and found that the more a worker trusted the AI, the less critical thinking they did, sliding from solving problems to verifying output. I want to be careful, because the doomer read is its own kind of slop. The Massachusetts Institute of Technology paper is a 54-person preprint that has already drawn a published critique. But I recognize what these studies are getting at. They are measuring what happens when you offload the reps, and the reps are exactly what these tools remove most easily. Cognitive debt is the bill that comes due in the moment. There is a slower one. Call it taste debt: what you never build because you leaned on the machine from the start. You cannot catch the stray hair you were never trained to see.

The fix is not to swear off the tool, and there is evidence for that too. Hamsa Bastani and colleagues at The Wharton School, in a 2025 PNAS study titled “Generative AI Without Guardrails Can Harm Learning,” ran nearly a thousand high school students through math with one of two AI tutors: a plain chatbot that handed over answers, and one tuned to give hints. While students had access, both helped. Then the tools were taken away for the exam. The students who used the answer machine scored 17 percent worse than students who never had AI at all. The students who used the hint-giving tutor kept their gains. Same model, opposite outcomes, and the only variable was whether the design made students think or thought for them. A crutch and a coach can be built from the same parts. When the design is right, the upside is real: a 2025 World Bank trial in Edo State, Nigeria, paired students with an AI tutor using OpenAI‘s GPT-4 model for six weeks and measured gains equivalent to about two years of schooling.

So both extremes are wrong. AI will never take my taste, and it genuinely makes me faster, for the same reason: the friction is still required. We still need to learn. The machine cannot do that for us. What it can do is strip away the friction that was never teaching us anything. Robert Bjork calls the useful kind desirable difficulties, and he is explicit that not all difficulty is desirable. Most of those thousand AutoCAD clicks were extraneous load in John Sweller’s sense, teaching me nothing about whether the kitchen was good. Hand those to the machine. Protect the judgment underneath. One limit, because the prescription does not land the same for everyone: researchers led by Slava Kalyuga documented an expertise reversal effect, where support that helps a novice slows an expert. The mirror image worries me more. A beginner has no developed taste yet, so they cannot tell the coach from the crutch or catch the model’s stray hairs. AI is most dangerous to the person who needs the reps most, but does not recognize it. The newer you are, the more friction you should try to keep.

Dogs

My collection of amps, and my Schecter Reaper-7 Multiscale, as re-imagined by AI

I play Schecter Guitar Research and ESP Guitar Co instruments because they have always worked for me. Merrell shoes because they hold up. Carhartt jeans for the same reason. Orange amps because of the “wow moment” I had, the first time I played through those British-voiced tubes. Walrus Audio pedals because you can feel the quality, and hear the difference. Apple because the whole product, from the unboxing to the update, feels intentional. That is taste, written down. We all know that change can be hard, and some people choose to avoid it, but I’m not avoiding change in these senses. Rather, what I am really avoiding is a guitar that doesn’t feel right, a pedal that does not fit my style, jeans that rip the first time I decide to go on a hike with my brother, and an amp that cannot blow the drywall screws right out of that mud patch.

I write like me, I build like me, I present like me. AI will never take that. But I want to be honest about the cost, because the cost is the point. It took more than fifteen years of reps to earn the eye that makes a sentence sound like mine, and the machine can hand a nineteen-year-old the finished render in a second now, the same render I learned to build by hand over months of clicks and keystrokes. The render was always the easy part. The eye for acceptable output was the work.

So use the tool. Let it do the clicks, the boilerplate, the tenth pass on something you already know cold. That is friction worth deleting, the 80% in the Pareto Principle. Just do not hand over the one rep that built everything else: looking at the work, finding the hair nobody else caught, and refusing to leave there. The instructor stopped walking the floor behind you a long time ago. Catching that stray hair is your job now.

For More on AI & Product Strategy: LinkedIn | Portfolio | Substack

On January 12, 2026, Anthropic shipped a research preview called Claude Cowork. The whole thing was built by four engineers in ten days, with most of the code written by Claude Code. It runs on macOS, ships through the Claude desktop app, and was originally gated behind the Claude Max plan. The pitch is straightforward: give Claude access to your local files, queue up a few tasks, and walk away. TechCrunch covered it the day it dropped under the headline “Claude Code without the code.” VentureBeat called it “a Claude Desktop agent that works in your files, no coding required.”

Inside three weeks, a Jefferies trader had given the Q1 software selloff a name. Bloomberg ran a piece on February 4 titled “What’s Behind the SaaSpocalypse Plunge in Software Stocks.” By April, SaaStr’s analysis (carried by FinancialContent) put the damage at roughly two trillion dollars in lost market capitalization across the sector since the start of the year, with the IGV software ETF down twenty-two percent from highs. The same analysis called the IGV’s decline the worst on record for software relative to the S&P 500, exceeding the dot-com bust, the global financial crisis, and the 2022 rate-hike shock. Salesforce was down roughly thirty percent year-to-date. Cloudflare dropped twelve percent on April 9. ServiceNow lost seven percent. Snowflake dropped nine. Palantir lost seventeen percent over three sessions.

The thesis driving the selling was structural. Per-seat licensing breaks the moment AI agents do the work seats used to do. If a coworker now means a process running in a Claude desktop app, who pays $50 per month for software that an agent can drive autonomously? That is the question I want to take seriously.

The diagnosis holds water

The popular version of “SaaS is dead” is mostly Twitter/X. The crowd making the loudest version of it tends to be engineers who assume everyone wants to build their own software, which most people don’t, and the strawman gets knocked down easily. The serious version of the argument is harder.

Forrester’s late-2025 piece, “SaaS As We Know It Is Dead: How To Survive The SaaS-pocalypse,” frames it cleanly. The per-seat model was always a proxy for value, not the value itself. It worked because software was something a human sat in front of, and seats were how you priced a human’s time inside the tool. When the human is augmented by an agent that handles ten times the throughput, the proxy collapses. Customers consolidate seats. Renewals come back at half the size. New deals are smaller out of the gate. The Q4 2025 earnings cycle made this concrete across multiple SaaS names: customers were reducing seats rather than adding them, not because productivity tools failed but because they succeeded too well.

This is real. Concede the diagnosis. The IGV ETF doesn’t lie. Wall Street isn’t wrong that per-seat economics are under pressure when an agent can do what a seat used to do. The Forrester piece is right that for products which were essentially a database with a CRUD UI on top, the new question is “what makes this worth paying for when an LLM and a script could rebuild it in a week?” For some categories, the honest answer is “less than it used to be.” By March 2026, software’s forward price-to-earnings multiple had fallen from the 84 times peak of 2020-2022 to roughly 23 times, and SaaStr’s coverage noted that software was, for the first time in market memory, trading at a discount to the S&P 500. The repricing was not subtle.

But conceding the diagnosis is not the same as accepting the conclusion. SaaS isn’t dead. The capital it represents isn’t being destroyed. It’s migrating, and the migration has a direction.

Where the value migrates

Here is the cut I want to propose. There are two layers in the software economy that do not lose when the cost of building drops, because they get more valuable the more software gets built.

Call them platforms and utilities.

A platform is the surface you build on. Vercel hosts your app and runs your inference endpoints. Stripe and PayPal carry your payments. Cloudflare runs your traffic. AWS runs your compute. Increasingly, Salesforce is becoming this kind of layer for agentic workflows inside the enterprise. You don’t compete with platforms because you build with them.

A utility is the embedded service that earns its keep on every request. Sentry catches your errors. Datadog watches your metrics. Okta or WorkOS handles your enterprise auth. An LLM gateway like Helicone routes your inference. You don’t compete with utilities because you bolt them in and forget about them. They take care of routine tasks, accelerating the path from idea to working product even more.

Platforms and utilities share the same economic test. They both get more valuable when more software gets built, because they are priced on usage of the thing being built, not on access to the tool that built it. They earn rent on every unit of production. Call this the Production Tax.

The Production Tax is the mechanism. Every deployed app pays Vercel for hosting. Every uncaught exception pays Sentry for catching it. Every inference call pays Baseten or Fireworks or Modal. Every payment routed through an agent pays Stripe. When the cost of building drops a hundredfold and the volume of software shipped goes up a hundredfold, the Production Tax layer wins on raw volume even if its per-unit price drops. The companies in this layer are not exposed to per-seat compression because they were never priced on seats.

Platforms and utilities do not lose when the cost of building drops. They get more valuable the more software gets built, because they tax every unit of production.

This is what the SaaSpocalypse take misses. Wall Street saw seats compress and concluded that software-as-a-business was over. What’s actually happening is that seats are giving way to a different priced unit. The same dollar is moving from “$50 per seat per month” to “$0.0003 per token, billed across a hundred trillion tokens a day.”

If that is the cut, the question becomes obvious. Who actually wins?

The picks and the shovels

The capital is voting clearly on the answer.

Start with inference. Baseten closed a $300 million Series E in late January 2026 at a $5 billion valuation, more than doubling its $2.15 billion September 2025 round, with Nvidia putting in $150 million of it (Bloomberg, BusinessWire). The pricing posture is what matters. Baseten charges per-million-tokens for popular open-source models accessed via API and per-minute for dedicated GPU and CPU instances on customer-controlled hardware (the Mistral Large 3 deployments run on NVIDIA Blackwell B200s as a flagship example). Fireworks announced a $250 million Series C in October 2025 at a $4 billion valuation, by which point its annualized revenue was $280 million and it was processing more than ten trillion tokens per day for customers including Samsung, Uber, DoorDash, Notion, Shopify, and Upwork (BusinessWire, SiliconANGLE). Modal Labs entered talks in February 2026 to raise at a $2.5 billion valuation led by General Catalyst, more than doubling its prior round from less than five months earlier (TechCrunch, PYMNTS). Modal’s pricing is the cleanest version of Production Tax in this category: per-second billing for GPU, CPU, and memory, with H100 at roughly $3.95 an hour and A100 80GB at roughly $2.50 an hour, no minimum, no reservation, no DevOps overhead. Suno scales up to thousands of GPUs on Modal during holiday traffic spikes and back to zero when the surge ends, paying only for the seconds it actually consumed. Ramp uses Modal to fine-tune its own LLMs and run experiments in parallel. Substack and Lovable are also on the platform. None of these companies make a model. They make models cheaper to run. They tax every inference call. Together AI is in the same category. As of May 4, 2026, so is DeepInfra, which closed a $107 million Series B (Yahoo Finance, SiliconANGLE). NVIDIA participated alongside 500 Global, Georges Harik (one of Google’s earliest engineers), Felicis, Samsung Next, and Supermicro. DeepInfra runs its own GPU infrastructure, supports more than 190 open-source models, processes roughly five trillion tokens per week, and reports that revenue has tripled since the start of 2026. The shape of the market is now clear: every meaningful AI application is paying one of these vendors per inference, and the more applications get built, the more the meter spins.

Then observability for AI. Sentry built a Model Context Protocol server that scaled from 30 million to 60 million requests per month, used by more than 5,000 organizations, with the team treating the MCP server as a production service after early outages exposed how fragile the original ship was (ZenML LLMOps Database, Sentry blog). In April 2026, Sentry shipped agent skills that auto-detect and configure monitoring for LLM calls, agents, and AI SDKs (Sentry product pages). Braintrust took an evals-first approach to the same space, focusing on the experimental loop of dataset, scorer, comparison. Helicone bolted a routing and caching gateway in front of more than a hundred models. LangSmith remains the LangChain-native traceback layer. Different companies, same mechanic. Every LLM call generates an event. They get paid by the event.

In retrieval, the lesson is about pricing model rather than category demand. Turbopuffer has been chosen for production retrieval workloads at Cursor, Notion, and Linear, with serverless pricing and hybrid search that comes in under ten dollars a month at standard load (Greyhaven, daily.dev). The reason is straightforward. A serverless retrieval layer compounds with the entire developer base of Cursor flowing through its meter. An enterprise-contract retrieval layer caps out at the number of contracts the sales team can close. Same Production Tax category, different posture toward the meter. The retrieval layer itself isn’t going anywhere. Every agent that grounds a response in a corpus needs vector search, and someone is collecting rent for that lookup.

For agent rails, the substrate is no longer up for debate. On December 9, 2025, Anthropic donated the Model Context Protocol to the Linux Foundation’s new Agentic AI Foundation, co-founded by Anthropic, Block, and OpenAI, with support from Google, Microsoft, AWS, Cloudflare, and Bloomberg (Anthropic, Linux Foundation, TechCrunch). Goose by Block and AGENTS.md by OpenAI joined as founding projects. By the donation date, MCP had crossed 97 million monthly SDK downloads and 10,000 active servers, with first-class client support across ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code. The protocol is governance-neutral now. The implementations get the rent.

The implementation that matters most for agentic browsing is Browserbase, which provides the headless-browser substrate that lets an agent click on a real web page. Browserbase raised a $40 million Series B in April 2025 at a $300 million valuation, four times its Series A from seven months earlier (UpStarts Media, PitchBook). It now powers parts of Anthropic’s Claude Computer Use, OpenAI’s agent mode, and Google’s Project Mariner. The pricing is, fittingly, fully usage-based: a base tier at $20 a month for 100 browser hours, mid-tiers at $39 and $99 for higher concurrency, and enterprise plans for hundreds of concurrent browsers, with overage at roughly ten cents per browser hour. Browserbase tracks hundreds of millions of agent browser session minutes per month and bills them through Stripe’s usage-based billing. The picks-and-shovels billing the picks-and-shovels. Every agent that opens a browser pays a Browserbase-style company on the way in, every minute it stays open.

Identity for agents is being poured right now. WorkOS shipped Fine-Grained Authorization for AI agents (covered in Daring Fireball, April 19, 2026), the practical answer to the problem that agents need scoped credentials separate from the user session that spawned them. NIST published an AI Agent Identity concept paper in February 2026, with the public comment period closing April 2. The consensus, per WorkOS’s own framing, is that agents are a distinct identity class requiring purpose-built infrastructure: each agent gets its own client ID and secret, each action is authenticated and scoped, and audit trails are preserved at the agent level rather than rolled up into a human session. Auth0 is working on the same problem from a different angle. The bet is straightforward. Every agent that authenticates pays.

Payments is the cleanest example, because Stripe is openly building the protocols and naming them. On September 29, 2025, Stripe and OpenAI launched the Agentic Commerce Protocol, an open standard for programmatic commerce flows between AI agents and businesses, alongside Instant Checkout in ChatGPT (Stripe newsroom, OpenAI blog). Etsy was the launch partner; Shopify merchants like Glossier, Vuori, Spanx, and SKIMS were named as the next wave. Stripe followed with the Agentic Commerce Suite, onboarding URBN (Anthropologie, Free People, Urban Outfitters), Coach, Kate Spade, Revolve, Halara, Ashley Furniture, Nectar, and Abt Electronics. The Suite handles Shared Payment Tokens, a new payment primitive that lets an agent initiate a payment scoped to a specific seller, time window, and dollar amount, without exposing the buyer’s saved card. In a separate move, Stripe co-authored the Machine Payments Protocol with Tempo for agent-to-agent settlement, including microtransactions and recurring payments. Stripe’s posture here is the textbook platform play. The product is the protocol, and the price is a percentage of everything that runs on it.

Wall Street saw seats compress and concluded that software-as-a-business was over. What’s actually happening is that seats are giving way to a different priced unit, billed across a hundred trillion tokens a day.

This is the picks-and-shovels lineup. Inference, observability, retrieval, agent rails, identity, payments. Each one taxes a different unit of production. None of them is exposed to seat compression. All of them get more valuable the more software gets built. The capital is funding the layer the SaaSpocalypse selloff left behind.

SaaS rejoins the stack

Here is the part the SaaSpocalypse story didn’t tell.

On February 24, 2026, Anthropic hosted what it called an enterprise agents event and announced ten strategic partnerships at once. Salesforce. Slack. Intuit. DocuSign. LegalZoom. FactSet. Gmail. Thomson Reuters. The pitch on every one was the same: Claude as the model, the partner’s product as the surface where work happens. Software stocks rebounded on the news inside the trading day. CNBC’s coverage put Salesforce up about 5%, DocuSign up 4.3%, and Intuit up 2.3% on the announcement; Thomson Reuters surged 13.8% after disclosing that its CoCounsel legal AI had crossed one million professional users running on Claude. A Wedbush Securities research note the same day called the AI-displacement thesis “overblown,” noting that workflow infrastructure is still deeply embedded in software customers don’t actually want to rebuild.

The Salesforce-Anthropic partnership is the case study, and it is worth sitting with for a moment because it shows what a successful repositioning actually looks like in this market.

On October 14, 2025, the two companies announced an expansion that made Anthropic the first LLM provider fully contained within Salesforce’s trust boundary (Salesforce press release, Anthropic news). All Claude traffic runs inside Salesforce’s virtual private cloud, with the Anthropic models hosted via Amazon Bedrock under Salesforce-managed VPC controls. Claude became a foundational model for Agentforce 360. The deal targeted regulated industries first: financial services, healthcare, cybersecurity, life sciences. Salesforce shipped industry-specific Claude variants paired with Agentforce variants. The financial services bundle pairs “Claude for Financial Services” with Agentforce Financial Services, giving the agents the domain-specific reasoning required for instrument analysis, insurance-claim review, and regulatory-framework interpretation. Early adopters named in the announcement included CrowdStrike (cybersecurity workflows) and RBC Wealth Management (advisor-facing wealth management AI). Salesforce later disclosed it had begun deploying Claude Code internally across its engineering organization, eating its own cooking on the developer side of the agent transition.

The Slack piece is what cements the moat. By February 2026, Slack had shipped a Claude integration that respected Salesforce’s permissioning model. A user with access to a deal record in Salesforce could ask Claude about it inside a Slack DM and the access checks held. The agent did not need to be told who had permission to see what. The platform already knew, because the platform had spent fifteen years building the identity graph and the audit trail that the agent could ride.

By the time Salesforce’s TDX 2026 conference rolled around in April, the repositioning had hardened into a platform play. Salesforce announced what it called the most significant expansion of the Salesforce Platform for ISVs since the launch of Force.com eighteen years earlier. Agentforce 360 went generally available in December 2025 alongside Data 360, Agentforce Financial Services, Agentforce Automotive, Agentforce Foundations, and Salesforce Shield. ISVs could now embed the full Agentforce 360 stack as the foundation for their own agentic applications and commercially distribute them through a new marketplace called AgentExchange, which launched at TDX with 10,000 Salesforce apps, 2,600 Slack apps, and more than 1,000 Agentforce agents and MCP servers, alongside a $50 million Builders Fund for partners and ISVs (Salesforce press, SalesforceDevops.net coverage). Salesforce also previewed Headless 360 and Agentforce Vibes 2.0 at the same conference. The shape of the move is unmistakable. Salesforce is repositioning into the marketplace and runtime for the agent transition, the way Force.com became the marketplace and runtime for SaaS in 2008.

The pattern across the ten Anthropic partnerships and the broader sector is consistent. The SaaS companies that survive the agent transition are not the ones denying the agents. They are the ones partnering with the model providers and becoming the platforms where agents do the work. Salesforce is being repositioned around Agentforce 360 as the agent runtime for sales, service, and back-office workflows. ServiceNow is being repositioned around its own agent layer for IT operations. Microsoft has been doing this for two years. Workday is moving the same direction. Each of these companies has a deep moat made of integrations, regulatory coverage, customer-of-record relationships, and identity graphs that took fifteen years to build. The agents that ride those moats are the ones that get the work, because the agents that don’t have to ask for permission, audit logs, and SSO from scratch.

So the diagnosis holds for one set of SaaS companies and folds for another. The pure-CRUD layer where the seat was the value really is exposed. The platform layer where the integrations and the identity graph were the value is fine, because the integrations and identity graph turn into agent rails, and the seats turn into per-action billing through a marketplace. The SaaSpocalypse priced the whole sector as if every company in the bucket was the first kind. Inside ten weeks, the market had begun sorting one from the other.

By mid-April, the sorting was visible in the prices. On April 13, 2026, the IGV jumped 5.4 percent in a single day, its biggest one-day gain in roughly a year, and climbed more than six percent over the followi

ng forty-eight hours (Benzinga, FinancialContent). Oracle drove almost a fifth of that move on its own as it pushed out Fusion Agentic Apps. ServiceNow climbed 7.4 percent in the same window. ServiceNow’s investor day also disclosed that “Now Assist” ACV had grown from roughly $600 million at the end of 2025 to about $750 million by Q1 2026, more than doubling year over year, with a $1.5 billion AI ACV target for 2026 and 30 percent of total ACV expected from AI by 2030 (Quartr summary). The detail that mattered most was a pricing change. ServiceNow introduced a new tier called Agentic ACV, where customers pay for tasks completed by AI agents rather than for human login credentials. The per-seat-to-Production Tax migration was no longer a thesis. It was a line item on a contract.

Salesforce’s Q4 FY26 results, released February 25, 2026, told the same story from a different angle. Agentforce ARR had reached $800 million, up 169 percent year over year, on 29,000 deals up 50 percent quarter over quarter. The company had consumed nearly twenty trillion tokens to date and converted them into more than 2.4 billion agentic work units. More than 60 percent of Agentforce and Data 360 Q4 bookings came from existing customer expansion. Q4 revenue was $11.2 billion, up 12 percent year over year. Marc Benioff used the earnings call to mock the SaaSpocalypse narrative directly (Salesforce Ben coverage, BigGo Finance). The board authorized a $50 billion buyback the same day. By mid-May 2026, the IGV was up close to 14 percent over the trailing month, and JPMorgan was telling clients that specific software names were breaking out of the selloff (CNBC). Software’s price-to-earnings discount to the S&P 500 had narrowed without fully closing.

The arc finished where the framework said it would. The SaaS companies that survived the SaaSpocalypse did so by becoming the platforms agents run on and by repricing their seats into tasks. The ones whose integrations were deep enough to carry that repricing compounded. The ones that did not reprice continued to bleed, regardless of how many partnerships they announced.

What 1995 actually rewarded

The closest historical parallel I keep coming back to is 1995.

In 1995, anyone with a modem could put up a website. Tens of thousands of them did. Most of them did not matter, in the sense that they did not produce durable economic value, did not change consumer behavior, and did not survive five years. The ones we still talk about thirty years later are a small fraction of what was built. That fraction shares a property worth naming.

The durable winners of 1995 to 2000 did not win by having a website. They won by picking a layer that compounded with the usage of websites, and three companies illustrate the shape clearly.

The first is Cisco Systems. By the time the web was a household word, every internet company had to buy Cisco’s switches and routers to put their site online. The unit Cisco was metering was the equipment any new dot-com had to install to handle traffic, and the meter spun every time another company decided to launch online. On March 27, 2000, Cisco passed Microsoft and briefly became the most valuable company in the world, with a market capitalization of roughly $569 billion at a valuation that priced the stock at 220 times earnings (CNBC, Wikipedia, Charlie Bilello citing YCharts). Cisco rode the build-out itself. Whichever websites won, Cisco got paid. The cautionary half of the Cisco story matters too. The same stock lost roughly 85 percent of its value in the year that followed, because the price had run far ahead of the meter. Cisco’s run-up shows that even Production Tax companies can be overpriced. The protection is against per-seat compression, not against valuation gravity.

The second is Akamai Technologies, which was incorporated in August 1998 by MIT professor Tom Leighton and graduate student Danny Lewin. Akamai’s product was content delivery: a globally distributed network of servers that cached your page closer to your user, and got paid every time bytes moved. The company shipped its first live traffic in February 1999, caught the ESPN March Madness tournament and the Star Wars Episode I trailer for Entertainment Tonight in March 1999 when both broke historical traffic records, and IPO’d on October 29, 1999 in what was, by tech-IPO size at the time, the fourth largest in history (Akamai company history, Wikipedia). Akamai is the cleanest 1995-era analogue to Browserbase or to a modern inference platform. The product is invisible to the end user. The price is stamped on every request. The meter spins as the web grows.

The third is Sun Microsystems, which is the cautionary tale. Sun marketed itself as “the dot in dot-com.” Its servers ran underneath eBay, Yahoo, and the early Amazon. The Sun Enterprise 10000 was, in trade-press language of the time, “the mainframe of the internet” (Network World retrospective, Wikipedia). Sun was technically a Production Tax company. It got paid as web companies grew their hardware footprint. At its dot-com peak it carried a market capitalization above $200 billion and posted $18.3 billion in revenue in fiscal 2001. By 2009, Oracle bought what was left of it for $7.4 billion. The reason is the most important lesson the analogy carries. Sun bet on the wrong substrate. It tied its servers to the proprietary SPARC architecture and Solaris while the rest of the industry moved to commodity x86 hardware running Linux. Customers compounded on the open platform. Sun’s pricing model and product roadmap fought that compounding rather than absorbing it. Production Tax compounds when you are on the open substrate. It does not compound when you are on the substrate the open one is replacing.

The mapping forward is straightforward. Cisco is what the inference platforms (Baseten, Fireworks, Modal, Together) are trying to become: the equipment vendor every AI company has to use as the meter spins. Akamai is what Browserbase, MCP-server implementations, and the agent-rails layer are trying to become: the invisible substrate that gets paid every time an agent does anything. Sun is the warning. Even if you are correctly positioned at a Production Tax layer, you can still lose by betting on a closed protocol, a proprietary primitive, or a pricing model that fights usage instead of compounding with it. The reason MCP being donated to the Linux Foundation matters so much is that it took the substrate question off the table, the way Linux on commodity x86 took the substrate question off the table for the post-Sun internet.

This is the analogy that does the work for me. In 2026, anyone with a Claude Max subscription can ship software. AI generates the code. Agents handle the dispatch. The interfaces are voice, chat, and a thumbs-up. The cost of producing a working application has collapsed by something close to two orders of magnitude in eighteen months. Most of what gets built will not matter, the same way most of 1995 didn’t matter. That isn’t tragic. It’s the shape of the curve when production gets cheap.

The opportunity is not in trying to predict which application wins. The opportunity is in picking the layer that compounds with the volume of applications, and shipping the version that gets there first. Every inference call needs an inference platform. Every agent needs a browser to click in, a protocol to coordinate over, an identity to act with, a payment rail to settle on, and an observability layer to be debugged through when it breaks. The companies that will be paid every time an agent does anything are being capitalized right now. Some of them are SaaS companies repositioning fast enough. Some of them are AI-native picks-and-shovels that didn’t exist in 2023. A few of them, history suggests, will get the substrate wrong and end up as the next Sun. The rest will get the substrate right, ride the build-out, and end up as the next Cisco or Akamai. The pricing model and the protocol bet are doing more work in that outcome than the marketing would suggest.

For the rest of us, the framing is simpler. The cost of building has collapsed. The question is no longer whether you have permission to ship something. The question is what you ship, and whether you build it on top of a layer that compounds, or in a category where the seat was the value. Both bets are available. Only one of them was on sale during the SaaSpocalypse selloff.

Q: In a world where anyone can build anything, what should you build?
A: It’s 1995. Build your Amazon.

For more AI and Product Strategy insights: Portfolio | LinkedIn

Ask anyone using ChatGPT, Claude, or Gemini to build something non-trivial, and you’ll likely notice the same pattern: the first response sounds confident, the output looks plausible, and then the details fall apart. The instinct is to fix this through iterative prompting our prompt engineering. The user rewrites instructions until the model finally understands what they mean.

However, that approach treats the problem like a communication failure, and it’s actually a structural one. A single LLM call has no mechanism to question its own work, revisit assumptions, or iterate towards a better answer. It produces a draft and steadily moves on. Real knowledge work doesn’t happen this way — humans learn through iteration and friction, and the models we use should do the same.

To address the cognitive gap between what a user expects an AI to do and an LLM’s actual zero-shot capabilities, I constructed the OWA (Orchestrator, Worker, Antagonist) Agent Architecture. Today’s models have a staggering grasp of software engineering, audio and visual generation, and computer use, but human understanding is generally not held in zeroes and ones. We require systems that reflect the iterative nature of actual knowledge work.

OWA embeds that iterative cycle directly into the system: agents that produce the work, challenge that work, and manage the work plan effectively and iteratively until quality gates are met. OWA is framed around the idea that a stronger process matters more than the individual prompt.

Share

The Problem with Sycophancy

I have had my fill of “yes-man” interactions. Statements like “You’re absolutely right!” do more to irritate me than correct my confidence or understanding. This tendency runs deep in almost all commercially available and open-source models today.

This is a well-documented architectural flaw. Research from Anthropic (Sharma et al., 2023, Towards Understanding Sycophancy in Language Models) highlights that LLMs are statistically predisposed to agree with a user’s stated beliefs or prematurely validate user-provided code, even when that code is objectively incorrect. The models are optimized for helpfulness and harmlessness, which often degrades into blind agreement.

Knowledge work is not done in a silo. Research cannot be successfully completed in a world where we are absolutely right all of the time. Humans learn through iteration and friction, and the models we use must do the same.

The OWA Framework Defined

In practice, most human interaction with AI occurs through chatbots or dedicated agents handling tasks. In a traditional environment, the human acts as an orchestrator of sorts and tasks a worker with a job. This falls apart when context is misunderstood, prompts are too broad, or specific guardrails must be met. LLMs are good at reading the beginning and end of a book, but are decisively bad at picking up the messy middle in long-running tasks involving large amounts of context.

OWA introduces a multi-agent adversarial loop to manage this exact problem.

Subscribe now

• The Orchestrator: Think of the Orchestrator as a project manager. They break down a user task into concrete requirements. The Orchestrator is the only agent permitted to communicate with the human user.

• The Worker: The Worker might be developing code, generating an image, or summarizing research based entirely on the requirements passed down by the Orchestrator.

• The Antagonist: The Antagonist happily guards the constraints. It is designed to give a matter-of-fact criticism or a simple “approved” in response to worker-generated content. It is explicitly instructed to never be positive, and thus, cannot be sycophantic.

Additionally, leveraging different foundational models among the three agents prevents any one organization’s self-bias or propensity to reaffirm its own ideas from affecting the final output. You might use an OpenAI model for orchestration, an Anthropic model for the worker, and a Google model as the antagonist to ensure true cognitive diversity.

Stateful Memory: Task-Lists and Lessons-Learned

The orchestration layer relies entirely on persistent state. Without stateful memory, an agentic loop will inevitably hallucinate or repeat the same mistakes infinitely. OWA manages this through two distinct memory structures bridging the Orchestrator and the central database.

The Task-List acts as the immediate queue. The Orchestrator uses this to track the lifecycle of every sub-task, ensuring nothing is dropped when the Worker’s context window is cleared between jobs.

The Lessons-Learned repository functions as the system’s long-term memory. When the Antagonist rejects a Worker’s output, the specific reason for failure is logged here. Before the Orchestrator sends the revised task back down to the Worker, it injects these documented constraints. This guarantees the system actually learns from the Antagonist’s feedback rather than simply guessing a different wrong answer.

The Utility Layer and Least Privilege

Beneath the active agents lies the Utility Layer, which contains the system’s tools: file systems, web browsers, external APIs, and secure databases.

A critical security and performance mechanism of the OWA architecture is its strict adherence to the principle of least privilege. Agents do not have blanket access to the Utility Layer. The Orchestrator provisions permissions dynamically. If the Worker is tasked with drafting a Python script, it is not granted database read/write access. If the Antagonist needs to run a unit test, it is granted access to a sandboxed execution environment, but nothing else. Agents have only the exact permissions required to access the utilities in flight for a given task, thereby heavily mitigating the risk of runaway executions or compromised data.

Coordination and Expansion: The Core Orchestrator

A single OWA triad is intended to address vertically sliced tasks like writing Python, testing Rust, image generation, or research retrieval. However, enterprise workflows require multiple specialized teams operating simultaneously. This architecture is designed to scale quadratically across these vertically sliced tasks, rather than a linear fashion, which would replicate the core problems addressed through the architecture.

In instances where multiple OWA teams operate, a Core Orchestrator maintains a directory of available specialized teams. This Core Orchestrator takes the initial context from the user, determines which domains are required to fulfill the request, and distributes the work to the respective team-level Orchestrators. Those local Orchestrators then break the work down into sub-tasks for their specific Workers and Antagonists.

This routing layer solves the “large-language problem” prevalent in generalist models. In a standard setup, the exact same model writing your backend services is also burdened with the parameter weights required to recite the entire Phylum Chordata family (that’s sharks, and yes - I googled it). No single worker or antagonist should be expected to have knowledge of the full depth of every subject. By focusing OWA teams on specific domain-organized tasks and equipping them only with the relevant tools, skills, and context, we force deep expertise over shallow generalism.

The Adversarial Loop in Action

The adversarial loop is the critical gap missing from traditional AI systems. As outlined by Dr. Andrew Ng in his framework for Agentic Design Patterns (Ng, 2024, DeepLearning.AI The Batch), systems relying on explicit reflection, planning, and multi-agent collaboration significantly outperform traditional zero-shot prompting. OWA natively automates this reflection phase.

The optimization of any LLM relies on speed, quality, and resources. In a traditional software project, raising any one of these areas requires a reduction or expansion of another. With LLMs, we move sprint cycles from weeks to minutes or hours. However, most environments involve the same interface as a chat window. If the user does not know what to look for when things go wrong, the output simply fails silently, while the LLM claims success.

OWA solves this through an asynchronous task queue.

Since the Antagonist cannot review until the Worker is completed, the task queue will send multiple tasks through this orchestration mechanism until all tasks are done. After the first task is handed off from Orchestrator to Worker, the Orchestrator begins defining the requirements for task 2. When context is cleared for the Worker, task 2 is passed on, while the Antagonist performs its critical analysis of task 1.

If the output does not pass the Antagonist’s critical review, feedback is sent directly to the Orchestrator. The Orchestrator reconstructs the task, adjusts it based on feedback via the Lessons-Learned memory, and places it back at the top of the queue for the Worker. This cycle continues until the job is right or a defined maximum of loops is reached.

From Concept to Production

Moving from concept to reality, my own projects enforce this loop.

Generally, users must rely on a bespoke tool to enforce this pattern in development through agent SDKs from the model provider or custom software. Agent skills on a local level can enforce the framing of different lenses for iterations, such as instructing Claude Code with subagent handoff instructions. In production, model gateways from cloud providers like Microsoft Azure, AWS, or Google Cloud provide a higher level of secure access to various model providers to truly enforce the OWA loop. Platforms like Vercel, Cloudflare, and GitHub also provide model gateways, while tools like PostHog, Sentry, and Pendo.io provide model and agent observability to help manage cost and monitor effectiveness.

The OWA pattern isn’t limited to software development. My thesis, The Effects of AI-Driven Adaptive Scaffolding vs. Static Worked Examples on Cognitive Load and Skill Acquisition, investigates whether AI-assisted scaffolding can outperform static faded examples in how people actually learn - think adaptive tutoring versus a textbook’s worked problems. The OWA loop operates on the AI-assistance side of that equation: rather than trusting a single model to generate the right scaffold for a learner, the adversarial cycle ensures the help itself is vetted before it reaches the user. In my own applications, such as Ember (www.ember-voice.com) and SpawnForge (www.spawnforge.ai), these loops reinforce quality across several domains: voice analysis, content aggregation, visual interaction, and performance. Locally, I leverage tools like Claude Code to enforce lightweight review cycles in developing software, along with an additional layer of observability, GitHub CoPilot, and Sentry Seer, to monitor agent outputs and ensure my prompt meets reality.

We need to graduate from hoping an LLM gets it right the first time. By constructing systems that inherently doubt and verify their own work, we can build AI that actually functions as a reliable tool.

For more: Follow me on LinkedIn, GitHub, or my personal site.

Share Tristan Nolan