Charles Watty

It’s easy to think of compliance as just something to “get through.” But sometimes, it’s exactly where the best product ideas hide. Take MFA. Strong authentication isn’t optional in most regulated markets. But here’s what happens when you lean into it: 💹 Push-based MFA works best through your native app. 💹 App installs give you a direct line to re-engage your users. 💹 Framing MFA as a trust feature (“We keep your account safe”) boosts opt-ins and reduces churn. Not every compliance requirement is just a cost of doing business — sometimes, it’s the unlock. This has come up for me more than once. Always good to swap ideas if you’ve been there too. #MFA #Compliance #Product #Growth

24

2 Comments

We want to make commerce better for everyone. That's why we're going beyond UCP working for Shopify on Agentic Storefronts. We're also open sourcing UCP Proxies as reference implementations for other commerce platforms. While it's DIY and advanced mode, it allows any merchant to participate in the new economy. Fork it, improve it, host it, enjoy it. Don't want to mess around with proxies but still want access to sell on ChatGPT, Gemini, Copilot and more? You can use Shopify's Agentic Plan without replatforming your Online Store. See: https://lnkd.in/gcgFeT8s

37

3 Comments

Synthetic ID fraud is skyrocketing (152% increase year-on-year for some financial products) The cause? Fraudsters are (ab)using AI to industrialise the creation of high-quality artificial identities. We're on it like a car bonnet.

15

Non-Deterministic AI Is the Fastest Way to Lose Money and Clients A reconciliation agent misallocated $340,000 across seventeen client accounts. Its job was to match incoming wire transfers to invoices by parsing bank descriptions and calling the accounting API. One Tuesday, a client paid three invoices in a single wire: "INV-2024-0892, INV-2024-0894, INV-2024-0901 - partial payment on 0901." The agent interpreted "partial payment" as applying to all three and distributed funds proportionally. Wrong. First two should have been paid in full. Credit hold triggered automatically. Shipment sat in warehouse for nine days. This is the failure pattern that kills agent deployments. Not formatting bugs - those get caught in week one. Destructive failures happen when agents process complex financial messages with implicit business rules humans parse unconsciously but language models interpret probabilistically. A wire description reading "Payment for Q3 services less $12,400 disputed amount per call with Sarah on 10/15" requires the agent to extract gross amount, identify the deduction, and flag the dispute. It works 97 times. On the 98th, it allocates full amount and closes the invoice. Dispute resurfaces three months later as a write-off. Tool-calling failures multiply when agents pass extracted data to systems that trust the input. A procurement agent created a PO for 200 units at volume-discount pricing because it missed a footnote saying prices valid only for orders over 500 units. Vendor rejected. Production line waited four days. Expedite shipping cost more than the agent saved in a quarter. The structural mistake: almost nobody writes sufficient acceptance scenarios for agents. Teams test happy path plus a few edge cases. Agent passes 10 invoices. Ship it. Missing from tests: invoice with handwritten correction, PO number in email body instead of attachment, two ship-to addresses, quote attached to six-month-old forwarded thread. These aren't edge cases - they're 30% of real volume. Agent reliability is not a model problem. It's a test coverage problem disguised as an AI problem. Same organization requiring 80% code coverage for billing system deploys an agent on the same data with zero acceptance criteria. Teams running agents successfully have test suites with hundreds of scenarios built from real production data, updated weekly with new failure cases. Every error is a missing test, not a fluke.

3

2 Comments

Fraud detection isn’t just about stopping bad actors. It’s about building trust at scale. I was recently invited to episode 2 of the "Fraud on the go" Podcast by Incognia to discuss how trust scores can help combat fraud and create smoother journeys for genuine users. For PMs, the real challenge is this balance: 🔹 Minimise risk without adding friction 🔹 Drive growth without compromising security If this resonates with you, check out the full conversation here 👇 https://lnkd.in/gBygtqVt

42

4 Comments

"The irony of 'general purpose' tooling is that it pushes the hard work onto users. Opinionated design is harder for us builders, we have to make calls, accept tradeoffs, be wrong sometimes, and face the music (ie. people on X who have opinions). But that’s exactly why it produces better products, you’re deep in the weeds doing the hard work every day." https://lnkd.in/deMZKDuu

1

Multi-Agent Systems - When One Agent Is Not Enough One agent writes code. Another reviews it. A third runs the tests. A fourth deploys. This is not a team of humans. This is a multi-agent system. Single agents hit a ceiling when tasks get complex. Context window limits, role confusion, and error compounding degrade quality as complexity increases. Specialization matters for AI agents just as much as it does for human teams. Three architectures that work: 1. Orchestrator Pattern: A manager agent delegates to specialists and synthesizes outputs (AutoGen, CrewAI) 2. Pipeline Pattern: Agents work sequentially, output becomes next input (LangGraph) 3. Debate Pattern: Multiple agents propose and critique solutions to converge on the best answer How agents communicate: structured messages with sender identity, task context, output format, and confidence level. The orchestrator maintains shared state. Protocols like Agent-to-Agent (A2A) standardize this communication. Real-world applications are already in production: - Software dev: Planner + Coder + Reviewer + Tester - Research: Search Agent + Summarizer + Fact-Checker - Customer support: Router + Specialist + Escalation Agent The coordination spectrum: Pipelines are easy (sequential, predictable). Dynamic delegation is hard. Emergent collaboration (agents self-organizing) is still research-stage. Frameworks: AutoGen (Microsoft), CrewAI, LangGraph, OpenAI Swarm. Each trades complexity for control differently. Have you built a multi-agent system? What worked and what did not? Share below. Follow for much more such content. #MultiAgent #AgenticAI #AutoGen #CrewAI #LangGraph

3

Stop hiding behind excuses for not letting LLMs touch your production codebase. “Legacy system.” “Too risky.” “Security concerns.” “Not production ready.” “Compliance.” If that’s still your narrative in 2026, you’re not being cautious — you’re being scared. And fear is not a strategy. The teams that win right now are not the ones debating whether LLMs are safe enough. They’re the ones building guardrails, shipping, learning, and moving. At Clerk.io, we run a big, complex, very real production system. Years of code. Deep domain logic. Edge cases everywhere. And yes — we let LLMs into that codebase. Not recklessly. Not blindly. But unapologetically. We use them to refactor. To explore architecture shifts. To accelerate feature development. To push ideas that would otherwise sit in a backlog for months. The result? We are shipping at a pace that would have been unrealistic just a year ago. And don’t watch what we release next year. Don’t watch what we release next quarter. Watch what we release in the coming weeks. This gap between teams who adopt LLM-native workflows and teams who hesitate is about to become permanent. There will be two types of companies: Those who figured out how to multiply their engineers. Those who protected their comfort zone. If you’re still saying “we can’t because…” — you’ve already started losing. Adapt. Build the guardrails. Train your engineers. Integrate the tools. Or step aside and watch me and other innovators redefine your industry. This isn’t gradual. It’s happening now. 🚀 #AI #LLM #EngineeringLeadership #BuildInPublic #FutureOfSoftware

10

I wrote about why product managers should step into the shoes of their customers and dispassionately figure out if they would buy their own product if they were a customer https://lnkd.in/gZDzuVFY

24

3 Comments

Good things come to those who think big, invent and simplify on behalf of the customer / user. That was going through my mind when I started reading about “MCP-UI”. While it’s tempting to rush and ship simple API wrappers as #MCP servers and just return verbose textual responses for agent to agent and agent to human interactions, that path ends up being just basic engineering 101. Next level innovation requires some deeper thinking and questions on what the interaction model between the agents and the humans should really be. That’s what folks like Ido Salomon and Liad Yosef, ⌞Samuel Path⌟ and others at Shopify have done. They have done a really cool job of extending the MCP protocol to return rich, interactive UI payloads back. Huge kudos! And, respect 🫡 This invention is a necessity for Shopify to provide rich commerce interactions that they eloquently explain in https://lnkd.in/gTp2hyeD I found the “intent based messaging system” that bubbles up the client side intent of actions like “Add to cart” for the backend agent to handle correctly to be a fascinating way to build stateful interactions. More on MCP-UI is here https://mcpui.dev/ I can envision the same pattern could be used in other places - for example, for a human subject matter expert sitting in #M365 interfaces receiving #ContactCenter channel engagements through #AI #Agents that require remote #CRM servers (such as #Dynamics) to be updated with multiple attributes for complex interactions. Can’t wait to see what the next application of this extension will be 😊 #AI #Agents #MCP #MCPUI #Dynamics

65

2 Comments

Strong signal this week. What felt fragmented for years is starting to align. After the Free Speech Coalition announcement yesterday, and following earlier commitments from Meta and Socure, seeing Persona, Incode and Veratad join the OpenAge Initiative confirms this is no longer marginal. Interoperable, privacy-preserving age credentials are moving from debate to real infrastructure. #AgeKeys are clearly emerging as the leading user-centric, privacy-preserving building block that can make interoperable age assurance work at internet scale.

17

1 Comment

We usually talk about deliverability in terms of technical setup: authentication, list hygiene, sender reputation. But Steven Lunniss from Cordial makes an interesting argument that deliverability really starts with intent. Send messages people actually want, and the inbox takes care of itself. It's a good reframe for anyone focused on email performance.

7

We'll learn a lot more about today's outages over the next couple of days as we start to get post mortems from the different platforms that were impacted. But it was interesting to see that traffic on our PNIs to GCP were not impacted uniformly with big traffic drops in LAX, ORD, SEA, SJC and DFW, but increases in traffic in FRA and AMS as customer services failed over. This highlights the important of proper DR planning with a minimum of multi-region and ideally multi-cloud.

825

21 Comments

Two months ago, I offloaded all daily ops to focus on agentic payments. Things are moving at light speed. I get impressed/inspired every day. Step 1: • get extremely hands-on Step 2: • force focus on what actually matters (the why) This shift is life-changing.

39

4 Comments

I've dealt with duplication for years in my work, and wrote up how I think you should approach it, and why you shouldn't panic when you find another team at your company is doing a similar thing: https://lnkd.in/gWfRKZwK

33

3 Comments

I've framed Agile Product Development as having 3 principles: Agility, Context, and Team Agency. So I'm drafting a rubric to "rate" an org, similar to DORA's 4-key-metrics rubric. Thoughts/comments/corrections? #productmanagement

2

2 Comments

🚨 Daily Harvest just made a $10M+ decision that most CEOs are afraid to make. They ditched their entire in-house software stack. For Shopify. Ordergroove. Klaviyo. Here's why this move is brilliant (and what it teaches us about scaling): Most growing companies fall into the "build everything" trap: → Custom software that requires constant maintenance → Internal teams stretched thin on tech debt → Resources tied up in non-core activities → Slower innovation cycles Daily Harvest realized something crucial: Your competitive advantage isn't your tech stack—it's what you DO with it. By migrating to best-in-class platforms, they: ✅ Free up engineering resources for product innovation ✅ Tap into proven, scalable infrastructure ✅ Access cutting-edge features without building them ✅ Reduce operational complexity and costs This isn't about being lazy. It's about being strategic. The smartest companies focus their internal resources on what makes them unique—not rebuilding what already exists. Daily Harvest can now invest in sustainable food innovation instead of maintaining payment systems. That's how you scale smart, not just fast. What's your take—when should companies build vs. buy their tech solutions?

5

A/B testing works best when it is not just a method, but a mindset. The article explores how testing helps teams shift from gut decisions to evidence-based thinking. Not just to improve conversions, but to drive clarity, alignment, and learning velocity. That landed. Because I have seen what happens when testing becomes routine: Teams start asking better questions Roadmap debates get grounded in data Risk feels lighter because learning becomes the default I still remember the day we hit the experiment limit in Google Optimize and I could not launch a new test. I mentioned it in a meeting as a small frustration. Our CEO gave me 50 bonus points on our internal reward system, not for the idea, but because we had five tests running at the same time. That moment said more than any metric could. If you are trying to build a test and learn culture, it is not about pushing more tests. It is about making learning part of how your team works.

2

One thing I’m noticing in digital identity is that the same problem gets described very differently depending on who you’re talking to. Fraud teams talk about risk exposure. Design teams talk about user friction. Growth teams talk about drop-off. Risk teams talk about appetite. Everyone is pointing at the same moment in the user journey, just through different lenses. What I’m learning is that product decisions don’t get harder because the problem is unclear. They get harder because alignment requires translating between these languages without losing intent. Still early in the space, but this translation skill already feels as important as any framework. Takeaway: great product work reduces uncertainty between teams, not just screens for users.

7

1 Comment