Nathan Clement

Most marketplaces eventually hit the same wall: Manual trust decisions → Hire more people → Build internal tools → Maintain legacy systems → Scale breaks I've seen this pattern up close. Tens of thousands of manual decisions daily. Teams working weekends, tickets piling up, burnout everywhere. Meanwhile? Customer experience nosedives. Commerce companies shouldn't have to build trust infrastructure from scratch. That's infrastructure's job. (This is why M8ven exists.)

20

I've been noticing the trend of "20x Companies" as YC puts it and it's surprising how productive and fast an increasingly tiny and microscopic team of people can outbuild larger industry titans. Large teams are getting slower. More layers, more coordination and approvals. At the same time, a 2-person team with a clear idea and a coding agent like Claude Code can now build what once required entire engineering orgs. The gap between idea and execution has collapsed. You don’t need massive headcount. You need focus, speed, and AI-augmented execution from now on. The barrier to competing has never been lower.

4

Troubleshooting production systems is one of the hardest problems in modern engineering. And two forces are making it dramatically more difficult: 1. The architecture of modern software is optimized for rapid change. Microservices, serverless, shared infrastructure, and multiple layers of abstraction are now the norm. AI generated code pushes the abstraction level even higher. It’s similar to how compilers hid assembly or how protobuf hid RPC mechanics. At this level of abstraction and pace of change, reasoning about system behavior becomes extremely hard for humans. 2. The information surface is massive. We generate enormous amounts of high dimensional telemetry across logs, metrics, traces, and events. Detecting anomalies in that scale of data is difficult. Inferring causality across those dimensions is even harder. Most failures aren’t caused by missing data but by the inability to connect the right signals at the right time. To move forward, we need systems that can reason at a scale humans cannot. Systems that understand opaque, fast-changing production environments, codebases, and telemetry. Systems that learn continuously from the data flowing through them. That is the future we are building, with AI SRE Agents, that apply reinforcement learning to adapt to dynamic production environments, generates dynamic and highly optimized code for large scale anomaly detection, are efficient operators for all the enterprise tools and continuously improving!

32

The way we build software is about to fundamentally shift, and most teams aren't ready for it. I'm watching something fascinating happen right now. Software development agents are moving from "nice to have" to "embedded in your actual workflow." Not in some separate tool you have to context-switch into. We're talking agents that live in your IDE, your Slack, your CI/CD pipeline, your project manager. Everywhere you already work. Here's what gets me: this isn't about replacing developers. It's about removing friction. When you spot a refactor that needs doing, you don't context-switch and spend an hour on it. You delegate it right there and keep shipping. When an incident hits at 2am, your team doesn't start from zero. The agent is already in the war room with them. The real competitive advantage isn't going to teams with the fanciest AI. It's going to teams that integrate this into their actual development rhythm without disrupting it. The ones who treat agents as part of their workflow, not a parallel process. For founders building technical products right now, this is worth paying attention to. Your development velocity just became a core business metric in ways it wasn't before. What's your biggest bottleneck in your current development process? The thing that takes time but doesn't actually require your creative problem-solving? That's probably your first candidate for this kind of delegation.

1

When we launched Thesys, one piece of feedback came up more than anything else: “Why can’t I use my own design system?” Fair question. Our generative layer was tightly coupled to the design system. That made it fast for us to ship… but hard for teams to bring their own components. So when we designed OpenUI, we made one deliberate architectural decision: Decouple the generative layer from the design system. Bring your own components. Any components. OpenUI handles the prompt generation, parsing, and rendering. Today that architecture gets its first real test. OpenUI now supports shadcn/ui. shadcn/ui has become one of the most elegant component systems in the React ecosystem. Composable. Accessible. Fully owned by the developer. It was one of the first things our community asked for. Perfect proof point for what OpenUI was built to do. Bring your own design system. Or add support for your favorite one. If you do, drop a GitHub link in the comments.

56

3 Comments

"Buy vs. build" is the wrong question for your software factory. The right one is architectural: can you swap any single piece without tearing up the rest? We're building one of these at SonderMind. The POC works, adoption is climbing, and I thought we had the shape right. Then Dex Horthy (HumanLayer) and Vaibhav Gupta (Boundary/BAML) spent an hour at a whiteboard on AI That Works and showed me we'd been treating one decision as four. A software factory has four layers: 1. Infrastructure. VMs, sandboxes, warm pools. 2. Dev environment. Toolchains, internal services, identity, preview URLs. 3. Harness. The agent loop: tools, skills, memory, testing. 4. Control plane. Dispatch, scheduling, review, governance, spend. You answer build-or-buy once per layer. And buying at one layer must never force you to buy the ones underneath it. Rent layer 3, don't marry it. The harness is the fastest-moving piece you have. Layer 2 is the one nobody prices in. No good abstraction exists for the dev environment: identity, the 50+ internal services a real app talks to, previews, the holes you punch so someone else's compute can reach your stack. Google and Facebook solved this for humans years ago with cloudtops. Agents are just new cloudtop users. Layer 2 is where our next month goes. And the seams between the layers? On Dex's diagram, half of them are drawn as "????". There is no standard interface between a control plane and a harness yet. That's the part worth arguing about. Last one: aim for 95%, not 100%. Boundary's bug triage sets a target accuracy on every step, and dedup at 60% ships because it still kills half the work. Most teams reach for 100%, miss, and drop the whole idea. Both links are in the first comment below: the full 70-minute episode, and the interactive field guide I built from it — the four layers, a build-or-buy matrix, the triage pipeline as a stepper. Who owns your dev environment: you, or a vendor? #ai-native #software-factory

7

1 Comment

Agents can plan, reason, and execute. But they still can’t pay. I was reading about Stripe's Machine Payments Protocol this week, We built an entire financial layer for humans. And now non-humans are trying to use it. That gap is exactly what the Machine Payments Protocol (MPP) is trying to close MPP basically bakes payments into HTTP itself. Request → “402 Payment Required” with a price → agent pays → retries with proof → server responds. It turns payments from a side quest into a first-class part of application logic. Why this matters if you build ML/agents/products: You can now design flows where your real user is not the human, but the agent acting on their behalf — and the payment infra actually keeps up. 🔹 Agents become real customers No more hacking around with headless browsers, fake emails, or manual top-ups. Agents can programmatically pay for APIs, models, data, or services they consume. 🔹 HTTP as a payment rail MPP standardizes 402 Payment Required with a formal Payment auth scheme, so “you need to pay” becomes a normal HTTP interaction, not a custom integration per vendor. 🔹 Microtransactions stop being a joke Per-request, per-session, per-token, per-email — the protocol is built for tiny, frequent payments that agents trigger autonomously. Think pay-per-browser-session, per-scrape, per-inference. 🔹 Works with existing money, not just crypto With Stripe, those agent payments can show up as stablecoins or good old fiat via cards and BNPL, all routed through Shared Payment Tokens (SPTs). 🔹 Merchants don’t need a “robot-only” stack For Stripe users, MPP payments land in the same Dashboard, balances, currencies, and payout schedules, using the same fraud, tax, and reporting stack as human payments. 🔹 Real use cases are already live Browserbase: agents spin up headless browsers and pay per session. PostalForm: agents pay to print and send physical mail. Prospect Butcher Co.: agents can order sandwiches for humans in NYC. 🔹 Plugs into the broader agentic stack MPP sits alongside things like Stripe’s Agentic Commerce Suite, Agentic Commerce Protocol (ACP), Shared Payment Tokens, and x402 — giving agents a full path from “discover product” to “pay and settle.” 🔹 Minimal code to get started On the Stripe side, it’s literally a PaymentIntent with payment_method_types: ['crypto'] and networks: ['tempo'] to accept MPP payments via Tempo. 🔹 New design constraint: agents as first-class users Product and infra teams now have to answer: “If 80% of our traffic is non-human, do our pricing, auth, and rate limits still make sense?” If you’re working on agents, APIs, or infra, this is one of those primitives that quietly unlocks a bunch of weird, new business models — especially “pay-as-you-go for everything” patterns. Curious to hear from you: Are you planning to let agents pay your product directly this year, or are you still in “humans-only checkout” mode? #AI #AgenticAI #Fintech #MachinePayments #Developers #MLEngineering

77

34 Comments

Five of the biggest names in software report earnings this week. Every call will come down to one exchange: we grew, and, what drove it? The market spent August repricing companies that answer that question with a narrative. This week we find out who has a record. The same referendum is coming for private companies. Your board is the analyst.

2

1 Comment

Had a conversation this week with a platform team that spent four months building a payment integration. Four months. They finally went live and immediately started hitting partial capture failures on split shipments. Their engineers had never dealt with auth-capture flows outside of a sandbox that auto-approves everything. Nobody on the team knew that partial captures have different rules depending on the processor, the card network, and sometimes even the merchant category code. This is the part that kills me. The code was clean. Architecture was solid. They had good engineers. But payments isn't just software - it's a regulatory and operational layer that punishes you for assumptions. You can't Google your way through settlement timing differences between Visa and Mastercard. You can't Stack Overflow your way past processor-specific refund windows on captured vs settled transactions. The gap isn't talent. It's domain knowledge. And domain knowledge in payments only comes from years of building inside it, debugging live transaction failures at 2am, and understanding why a decline code 05 means something completely different depending on which issuer sent it. If your engineering team hasn't lived in payments, they're going to learn - but they'll learn on your merchants' money. #payments #paymentinfrastructure #fintech

2