The Chaos Is the Point
When the most talked-about AI consumer product of early 2025 turned out to be a semi-rogue open source project that went through three name changes in three days, communicated over WhatsApp, and stored its memory by scribbling notes into a markdown file like the protagonist of Memento, Bret Taylor found it delightful.
"People in our circles have been talking about AI agents for consumer use and all these fancy computer using agents and instead you're chatting over WhatsApp with a thing on a Mac mini that is, you know, mildly unhinged and insecure. It's just fascinating."
Taylor — co-creator of Google Maps, inventor of the Facebook Like button, former co-CEO of Salesforce, chairman of the OpenAI board, and now founder and CEO of Sierra — sat down with Stripe co-founder John Collison for a wide-ranging conversation on the podcast Cheeky Pint. What emerged was a masterclass in how one of Silicon Valley's sharpest minds is thinking about AI agents, the future of software, outcome-based business models, and why the janky, markdown-scribbling approach to AI memory might actually be more profound than it appears.
Why Markdown Beats MCP (For Now)
The conversation opened with an observation that captures the strange state of consumer AI in 2026: you can use super-polished mainstream apps from Google or OpenAI that have essentially no memory, or you can use a wildly insecure hobbyist project that almost remembers things by writing notes to itself in a file. Taylor argued that this isn't just a hack — it might be pointing toward something fundamental.
His reasoning starts with why coding agents have improved so dramatically. Code repositories have specific qualities that make them ideal for AI: all the context is in one place, in textual files, with built-in feedback loops like compiler errors, unit tests, and formal version histories.
"It's almost designed for a robot and you can actually... it's almost like we as engineers are sort of like have always modeled ourselves after robots and now we can actually, you know, fully realize that vision."
The markdown memory file, Taylor suggested, is actually an attempt to turn the rest of life into something that looks like a code repository — a structured, text-based context store that AI agents already know how to navigate. Vector databases offer random access, but that's not how real memory works. The messiness of a markdown file, with its mix of context and sequential notes, may actually produce a more useful agent than fancier approaches.
This insight led Taylor to express growing skepticism about MCP (Model Context Protocol) as a meaningful part of the future architecture. The vision of specialized sub-agents connected through formal protocols looks elegant on a whiteboard, he argued, but strips context from the orchestrating agent in ways that make the experience feel robotic.
"What ends up happening in these multi-agent systems is you stuff all the context in the sub-agents and the one on top has no ability to actually like not sound robotic. And then in contrast, you look at something like Open Cloud, it's just a bunch of Markdown files and it's... the memory kind of feels right."
AI Agents Are Already Talking to Each Other — Over the Phone, in English
Perhaps the most surprising anecdote of the conversation came from Taylor's work at Sierra in healthcare. The company powers AI agents for health insurance payers like Cigna and Blue Cross Blue Shield, healthcare providers like Sutter Health, and revenue cycle management companies. These organizations all need to call each other — providers calling payers about procedures, revenue cycle companies calling to ensure payment.
Sierra has already had AI agents on both sides of phone calls — one agent making the call, another picking it up. And how do they communicate? Not through some elegant machine-to-machine protocol, but through English, over the publicly switched telephone network.
"We've done English over the publicly switched telephone networks... And it goes... it sort of reinforces Dario's point which is, you can engineer all these fancy protocols, but the rails that are already there already exist."
The image is both absurd and illuminating: two AI agents speaking English to each other over century-old telephone infrastructure, while Silicon Valley debates the future of agent-to-agent protocols. It underscores Taylor's broader point that backward compatibility with existing systems — whether that's Unix tools, telephone networks, or plain text files — often beats elegant new architectures.
The Death of the Dashboard (and the Rise of the Harness)
Taylor painted a provocative picture of how software will be consumed in the future. Today, SaaS applications have two interfaces: a web application for humans (dashboards, forms, buttons) and APIs for machines. Taylor believes a third layer is emerging that may become the most important: the agent harness.
A harness, in Taylor's framing, is more than an API. It's the full context an AI agent needs to extract maximum value from a product — the documentation, the skills, the rules, the institutional knowledge. He used Stripe as an example:
"Imagine the person who's the greatest Stripe expert who knows how to extract the most value from their Stripe account. That's the harness. Not the API. That's just the button you click."
This leads to a fascinating possibility: the quality of a company's agent harness might become a key competitive differentiator. Two years from now, Taylor suggested, companies might evaluate their software vendors partly on how well they work with the AI agents managing their operations.
Collison noted that many SaaS dashboards have switches and settings with no corresponding API, and the entire industry is collectively discovering this gap. Taylor's response was characteristically pragmatic: in the agent-first world, you might not need beautiful UI design. A dashboard that looks like "a Russian submarine of switches" would be fine, as long as the harness describes when to use each one.
Sierra's Rocket Ship — and Why Customer Service Is Just the Beginning
When the conversation turned to Sierra's business, the numbers spoke for themselves. The company reached $100 million in annual recurring revenue in seven quarters and $150 million in eight quarters, with Taylor noting they were around $165 million just one month into their next quarter.
But the metrics only tell part of the story. Taylor's vision for Sierra extends far beyond traditional customer service. He pointed to Rocket Mortgage as an exemplar: through their ownership of Redfin and acquisition of mortgage servicer Mr. Cooper, Rocket is using Sierra's AI agents for everything from home searching to mortgage origination to loan servicing.
"If you go to Redfin, you can search for a home using the AI agent. If you go to rocket.com, you can originate a mortgage with an AI agent and you can service that mortgage."
This represents a fundamental shift from customer service as a cost center to the AI agent as a company's primary digital interface. Taylor drew the analogy to 1994 and the early web:
"I think a company's AI agent will be the vast majority of their digital interactions. And I think digital has come to include the telephone, and that's sort of a big shift."
The economics are staggering. A human-handled customer service call can cost $10 to $20. AI agents can bring that to 10 to 20 cents, eventually one to two cents. But the most interesting effects aren't just cost savings. One retailer saw total conversation volume increase nearly as much as their cost savings — a kind of Jevons paradox for customer service. People actually wanted to talk to the AI agent because it was so much better than the old chatbots.
"They ended up keeping, their cost didn't really go down, but their volume of customer conversations went up, you know, two or 3X. And the CEO was incredibly happy about it. They were like, 'We're now actually listening to our clients.'"
The Supervisor Model: How Sierra Keeps AI Agents From Going Rogue
For companies like Cigna or SoFi (where Sierra raised net promoter scores by 33 points), trust is everything. Taylor explained Sierra's approach to reliability: a "constellation of models" where supervisor agents inspect the reasoning of the primary agent.
"If you imagine a reasoning system is right 90% of the time... and then you have a supervisor that's right 90% of the time, if you chain them together, you get 99% effectiveness."
One counterintuitive challenge: it's actually harder to ground AI agents for well-known brands than obscure ones. When an LLM has been trained on vast amounts of public information about a company, it's tempted to answer from its training data rather than looking up the actual current policy. The model essentially says, "I got this," when it really doesn't.
Taylor also shared a charming example of how the innate knowledge of LLMs creates unexpected value. Sonos, one of Sierra's clients, uses an AI agent for customer support. Most Sonos problems turn out to be wifi issues, not speaker issues. The AI agent is effective not just because it knows Sonos products inside and out, but because large language models have encountered every possible wifi problem in their training data.
"Why does the Sonos AI work so effectively? Well, it knows a lot about wifi in addition to all the Sonos things."
Outcome-Based Pricing: The AdWords Moment for Software
Taylor is a vocal advocate for outcome-based pricing — a model he sees as fundamentally different from usage-based pricing. Under Sierra's model, customers pay a pre-negotiated rate when the AI agent successfully resolves a case without human intervention. If the agent has to escalate to a person, it's free.
"Reducing your token utilization for the same outcomes is your problem, not your customers. And that's a great incentive to just drive more efficiencies over time."
He drew a sharp distinction between usage-based and outcome-based approaches. Token consumption, he argued, is like lines of code — not necessarily correlated with value. He referenced the famous story from Apple's Folklore.org where an engineer reported negative lines of code after refactoring. The analogy is clear: what matters is the outcome, not the compute consumed to get there.
Taylor compared this shift to the advertising industry's move from impression-based to cost-per-click pricing — a change that made the entire market more efficient and more valuable.
The Software Industry Reckoning
Public markets have marked down the software sector by 20-30%, and Taylor thinks the reaction is rational, if somewhat overblown. His analysis is nuanced: there has never been more uncertainty in the software market, and if you're investing in a portfolio of software companies, a discount makes sense. But it's an indictment of the sector, not necessarily of individual companies.
The deeper question, Taylor argued, is whether AI agents will upend the 30-year-old structure of enterprise software, where value concentrated in systems of record organized by department — Salesforce for sales, SAP for finance, ServiceNow for IT. If AI agents are actually performing valuable labor, the encoded, well-optimized processes they embody might become more valuable than the databases they draw from.
"Agents are, to some degree, a system of record of a process. Of generating a lead or auditing your financials or reviewing a contract or whatever it might be. And I don't think we've ever had a piece of software like that."
His framework: the closer a system of record is to being literally a ledger (like an ERP system), the more durable its value. The closer it is to being a system of engagement, the more vulnerable it becomes.
Building to Throw Away
Perhaps the most revealing window into Taylor's thinking came when he described the strange experience of building a company in the current AI landscape. Sierra routinely builds technology it plans to discard as model capabilities catch up — like their best-in-market Cantonese voice support, which Taylor is virtually certain will be commoditized within three years.
"You can't have the luxury of waiting for all the models to catch up with your aspirations. But you know they will. So you have to have the best technology and have to be comfortable with throwing it out."
When Collison asked the pointed question — "Isn't Sierra itself kind of short AGI?" — Taylor didn't flinch.
"The fog of war in the software industry is pretty thick right now."
But he expressed deep conviction in the applied AI market, arguing that most companies want solutions, not models or raw software. He made a striking claim:
"If we paused model development, we'd still have trillions of dollars of economic value that have yet to be realized."
The Process Revolution
Taylor's most provocative argument was about how companies need to reorganize themselves to capture AI's benefits. The atomic unit of AI productivity, he argued, is a process, not a person. Companies are organized by departments — legal, procurement, finance, IT — but the real value of AI shows up in optimizing end-to-end processes that cut across those departments.
Take something as mundane as onboarding a new supplier: it involves legal contracts, procurement negotiation, IT system integration, and business sponsorship. If the median onboarding time is 17 days, AI could potentially compress it to 17 hours — but only if someone owns the entire process, not just their department's piece of it.
"I think part of the reason why I think it's been slow to get the productivity enhancement is we sort of ship our org charts as companies, naturally. There's not usually a person responsible for that process."
He was equally pointed about companies that think buying CoPilot licenses for every employee constitutes an AI strategy. Real transformation requires thinking from first principles about which business processes have heavy digital workflows and can benefit most from AI — then actually assigning accountability to drive those improvements.
Conclusion: The Recipes Are Changing
Throughout the conversation, Taylor returned to a metaphor that captures his worldview with elegant simplicity: AI has changed the ingredients available, and now there are far more recipes to cook. The most impressive companies won't be the ones that adopt AI as a cost-cutting tool and pass the savings to shareholders ("unless you have a monopoly," he laughed). They'll be the ones that recognize AI adoption as an imperative — not a competitive advantage — and use the window of disruption to fundamentally reimagine their businesses.
The state of AI in 2026 may be captured perfectly by the image of two AI agents speaking English to each other over phone lines built a century ago, while a hobbyist project on a Mac Mini writes markdown notes to itself like a digital amnesiac. It's messy, it's chaotic, and it's wildly productive. For Taylor, that's not a bug — it's a feature. The elegant architectures will come later. Right now, the companies that win will be the ones comfortable building in the chaos and throwing away what they build tomorrow.
"We've now, like, changed the ingredients available and we have a lot more recipes we can cook. And I think that's very exciting."