Diego Granados

If your team is deploying AI agents, you need agent observability. Monte Carlo's agent observability makes it easy for data + AI teams to gain visibility into the four critical pillars that determine whether agents can operate reliably in production: context, performance, behavior, and outputs. Check out our newest features, including Agent Metric Monitors, Agent Trajectory Monitors, Agent Evaluation Monitors, OpenTelemetry deployment in AWS, and more: https://lnkd.in/e6zjxBQR #AgentObservability

11

The Two Spirals of Product Development Many of the ideas here are inspired by Marty Cagan (Inspired, Empowered), Reed Hastings and Erin Meyer (No Rules Rules), and other great product leaders. The spiral concept itself comes from my own observations of how teams succeed or stall in delivering value. Product development operates in a spiral — either upward or downward. In the positive spiral, teams: 1. Balance quick wins, small bets, and big bets 2. Deliver consistent customer value 3. Earn trust 4. Gain autonomy 5. Deliver even better outcomes In the negative spiral, teams: 1. Miss repeatedly 2. Accumulate tech debt 3. Lose trust 4. Get more gates, more dates, more process 5. Slow down — miss more Trust follows outcomes. Autonomy follows trust. What makes the difference? Small, context-rich teams with high talent density where Product, Tech, and UX truly operate as one Team to solve customer problems in ways that work for the business. Not PRD → Tech → UX. Not “agile” ceremonies masking waterfall. Real shared context: - PM ensures the problem is worth solving - Tech ensures the solution is feasible (including cost and performance) - UX ensures it’s usable and minimal design Great developers are critical in this. They don’t just build — they think in terms of problems, propose solutions, and take ownership. They balance value, feasibility, and usability, care about quality, and thrive when context is shared and outcomes matter more than instructions. Steve Jobs once mentioned Apple’s strength in decomposing problems well and trusting other teams to deliver their part. That only works with clear (product, tech and people) interfaces and small, empowered teams. When a team introduces 2 months of tech debt to avoid a 1-week delay in hitting a date, in 99.9% of the cases that’s not speed! AI amplifies this dynamic. It increases iteration speed — which makes trust, judgment, and team empowerment even more critical. The goal for Product Teams: get to the positive spiral :)

23

2 Comments

Let’s be honest: Most “AI in product” conversations are still about chatbots, copilots, or shiny new features. But the real transformation is happening behind the scenes, in how teams decide what to build next. Here’s how high-performing product teams are already using AI to sharpen decision-making: > Idea triage with LLMs. Teams feed customer feedback, support tickets, and sales notes into models that cluster themes automatically, surfacing the top pain points in hours, not weeks. > Scenario modeling. PMs are using small agent networks to simulate user behavior, pricing changes, or feature interactions before launch. (Think “digital twins” for your roadmap.) > Impact forecasting. AI models trained on historical product data now predict likely engagement lift, churn impact, or cost-to-serve before resources are committed. > Faster retrospectives. LLMs summarize post-launch data into executive summaries that focus on insight — not slide formatting. The secret isn’t that AI is replacing PM intuition. It’s that it’s removing the noise so teams can focus on the signal. If you could automate one part of your roadmap process tomorrow, what would it be? Here’s a recent example of AI in product decision-making from the field: https://lnkd.in/g4sMUyrv — “Multi-Agent Collaboration for Complex Decision Support” #AI #ProductManagement #Roadmapping #Innovation #FutureOfWork #DataDriven #Leadership #LLMs #Strategy

4

Everyone’s racing to become “AI-centric”. AI-centric products. AI-centric orgs. Even AI-centric companies. But in the real work, where priorities compete and reviews come fast, I’ve been hearing a quieter theme from peers: ROI is vague, mismeasured, or expected way too early. Some teams are asked to show impact just one month after implementation. No baseline. No clean comparison. Just pressure. I’ve even seen “AI roadmaps” where teams brainstorm how to force AI into their product - not because it’s needed, but because it’s expected. That creates its own risks. AI can generate powerful insights, but without the right context, it’s easy to mistake: 1. Noise for signal 2. Insight for truth 3. Garbage for gold I recently saw a tool learning from Slack channel conversations. Maybe it will surface something useful - but my first reaction was skepticism. Slack is full of side tangents, missing context, half-decisions. I’d love to be proven wrong. To be clear, I’m not skeptical of AI. I’m excited. It’s lowering the barrier to build in ways that were unthinkable even a few years ago. (I’ve been exploring product ideas of my own, more on that another time.) But this is the tension I keep circling: How do we stay open to long-term potential without overengineering today’s problems? One question I’ve started using in AI reviews: “Would a rule solve this just as well?” Sometimes the answer is yes. And sometimes, it’s a chance to build real leverage. Would love to hear how others are thinking about this in real systems.

9

4 Comments

The last 30 hours at an AI summit (in ATL) have stretched me in fun, exciting, and frustrating ways. On about four hours of sleep, which somehow made it even more interesting with Hoop Somuah, Sedem Tay-Agbozo, Peter Somuah and a few other wickedly smart, AI focused individuals have been deep in agentic AI, spec-driven development, and what it actually means to build with agents instead of defaulting to the workflows we’ve used for years. At one point during the summit, I decided to go all in. I barely touched my favorite traditional IDE and leaned fully into agent-driven building. It was enough for me to cancel my IntelliJ license, which I’ve had since 2007. That felt symbolic. I can now say I’m a Claude Code convert, and it has meaningfully changed how I approach and will approach software. We didn’t just stay in software either. We played with hardware. Connected two machines over USB. Used Claude to help install a fresh operating system onto a blank machine. Broke things. Fixed things. Pushed boundaries. There’s something different about watching an agent reason through system-level setup instead of just generating application code. It makes the shift tangible. The biggest mental change for me has been this: I spend far less time reading code and far more time evaluating output.. Honestly, I don't know if that is a good thing or not, but it is part of the journey I have decided to go on. Steve Yegge argues that we should care more about what the system produces than manually inspecting every line generated by an LLM. When you internalize that, the economics of development look different. Once the spec is clear, generating code is no longer the constraint. Clarity is. Judgment is. Validation, integration, architecture, and system design are. The leverage is moving from typing to thinking. From implementation to orchestration. From individual productivity to parallel throughput across intelligent agents. This is not about coding faster inside the same container. It’s about redesigning the container itself. How we define intent. How we evaluate output. How we structure teams. How we build guardrails. We also learned quickly that without constraints, things can go sideways. Broad machine access without guardrails is risky. Vague tasks can burn through tokens fast. It gets expensive quickly. Tools like OpenClaw are incredibly powerful, but they also make you pause and think about the security implications. Are we really ready to hand over credentials and system access without strong governance? The technology is accelerating. Discipline has to accelerate with it. If we embrace this shift seriously, the bottleneck is no longer code production. It’s clarity of intent and the quality of judgment applied to intelligent systems. That changes what engineering excellence means. Agents can write code. The real issue is whether we’re ready to rethink our development mindset. #AgenticAI #SoftwareEngineering #DeveloperProductivity #FutureOfSoftware

238

30 Comments

rqlite -- and the way I develop it -- is fundamentally changing. LLMs are in the loop -- from ideation to Pull Request. https://lnkd.in/ejwXGddR

27

1 Comment

Today we're open-sourcing StyleMe on Github My 9-year-old co-founder's first reaction: "Why are we giving it away?" I tried explaining open source to her and here's what actually landed: In the age of AI, producing code is essentially free. We built this entire app using #Claude Code over spring break. The code isn't the moat. The ideas, the design decisions, the 8 spec documents we wrote before touching a line of code — that's what made it work. If the community can take what we built and make it even better, we all benefit. Better prompts. Better UX. Better outfit suggestions. More wardrobe items tagged correctly. That's the deal. She's 75% convinced. The other 25% still thinks we should charge for it. Honestly, her business instincts might be sharper than mine. What you'll find in the repo: · Full Next.js + Convex + Claude API stack · 7 spec docs — PRD, tech spec, UX wireframes, prompt templates, build order, and more · A swipe-based preference learning system · Everything we learned about prompting Claude to not suggest two sweaters in one outfit

5

Elena Verna, featured on The PM Brief today, highlights a shift many of us are already experiencing in product work. As AI reshapes how products are built and shipped, the question is less about tools and more about how our skills evolve alongside them. In practice, this often means working closer to systems, iteration speed, and new forms of product feedback. It’s an exciting time to be in product, and a reminder that continuous learning is part of the job!

10

1 Comment

Had a blast presenting at #CCW Berlin today with Paul Biggs 👑. We talked about how to put a leash on your AI-Agents operations before they spiral out of control. The culprit? Copy-paste scaling. One agent becomes ten. Ten becomes fifty. Suddenly you're herding cats. The fix? Agent Composition — one agent, every market, every language, every channel, every variation. Link to learn more in comments 👇 #CCW #Parloa

107

7 Comments

The Quiet Enshittification of a Beloved App Apple, a company celebrated for human-centered design, ships a music app that is unstable in behavior, opaque in navigation, hostile to power users, and persistently redirects intent away from the user’s own library toward algorithmic consumption. Core library management is subordinated to promotional surfaces. That doesn’t happen by oversight. It happens when incentives shift. What was once a personal media tool is now a content distribution surface. Retrieval loses to discovery. Stability loses to experimentation. User continuity loses to engagement optimization. The result feels confusing not because it’s poorly executed, but because it’s executing on goals that are no longer aligned with the user. That’s a subtle form of enshittification: not ads everywhere, not obvious degradation—just a steady erosion of agency in favor of platform objectives. Good design didn’t disappear. It was pointed somewhere else.

8

1 Comment

Agentic AI — Before LLMs Agentic AI isn’t just about tools or autonomy. Some of the most effective agentic systems operate inside human interactions. As LLMs became mainstream, I started reflecting on earlier work I led at Amazon Alexa—and it became clear that we had already built what today would be called an agentic AI system. At Alexa, I worked on voice frustration detection—not to make Alexa more verbose, but to help her recover broken conversations in real time. Looking back, it clearly fits today’s definition of agentic AI: • A clear goal: detect frustration and restore engagement • Perception: tone, prosody, and interaction signals • Reasoning: assessing user state and severity • Decisioning: choosing the right recovery action • Memory: recent context and failures • Evaluation: did the intervention actually help? • Guardrails: strong privacy, safety, and trust boundaries Alexa wasn’t just generating responses. She was deciding when and how to act. This system existed before LLMs were part of the stack. What LLMs unlock today is more flexible reasoning—but the agent loop already existed. The real challenge of agentic AI isn’t maximizing autonomy. It’s designing systems that know when to act, how to act, and when not to. That’s where trust is built—and where real product value emerges

16

2 Comments

🚀 𝗔𝗱𝘃𝗮𝗻𝗰𝗶𝗻𝗴 𝗪𝗼𝗿𝗹𝗱 𝗠𝗼𝗱𝗲𝗹𝘀 𝘄𝗶𝘁𝗵 𝗩‐𝗝𝗘𝗣𝗔 𝟮 - Excited to share the latest milestone from one of the research teams I work with at Meta FAIR Core Learning & Reasoning. V‑JEPA 2 is a world model to understand physical reality, anticipate outcomes, and plan efficient strategies. 𝗪𝗲, 𝗵𝘂𝗺𝗮𝗻𝘀 𝘂𝘀𝗲 𝗼𝘂𝗿 𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝘁𝗼 𝗽𝗿𝗲𝗱𝗶𝗰𝘁 𝗵𝗼𝘄 𝘁𝗵𝗲 𝘄𝗼𝗿𝗹𝗱 𝘄𝗶𝗹𝗹 𝗿𝗲𝘀𝗽𝗼𝗻𝗱 to our actions—or the actions of others— all the time, especially when planning what actions to take and how to best navigate new situations. 𝗩‐𝗝𝗘𝗣𝗔 𝟮 𝗶𝘀 𝗮 𝘀𝗲𝗹𝗳-𝘀𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗲𝗱 𝘄𝗼𝗿𝗹𝗱 𝗺𝗼𝗱𝗲𝗹 that reasons in learnt latent space - not in token space like LLMs. The encoder and predictor are pre-trained through self-supervised learning from visual data, leveraging abundant natural videos (more than 1 million hours) from diverse sources to bootstrap physical world understanding and prediction. 𝗔𝘀 Yann LeCun 𝗵𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝗲𝗱 𝘁𝗼𝗱𝗮𝘆 𝗮𝘁 𝗩𝗶𝘃𝗮𝗧𝗲𝗰𝗵, this represents a fundamental shift toward AI systems that can reason, plan, and act through physical world models. To reach advanced machine intelligence, AI must go beyond perception and understand how the physical world works—anticipating dynamics, causality, and consequences. V‑JEPA 2 does just that. 𝗥𝗼𝗯𝗼𝘁𝗶𝗰𝘀 a𝘀 𝗮 𝗽𝗼𝘄𝗲𝗿𝗳𝘂𝗹 𝗲𝘅𝗮𝗺𝗽𝗹𝗲: with just 62 hours of robot interaction and no task-specific training, V‑JEPA 2 enables zero-shot planning and achieves up to 80% task success, reducing latency from minutes to seconds. It’s a leap toward practical, adaptable robots that can operate in dynamic, real-world environments. 𝗧𝗵𝗲 𝘁𝗲𝗰𝗵 𝗰𝗼𝗺𝗺𝘂𝗻𝗶𝘁𝘆 𝗶𝘀 𝘁𝗮𝗸𝗶𝗻𝗴 𝗻𝗼𝘁𝗲 and on Hacker News, one commenter summed it up well: “𝘐𝘵'𝘴 𝘪𝘮𝘱𝘰𝘳𝘵𝘢𝘯𝘵 𝘵𝘰 𝘬𝘦𝘦𝘱 𝘴𝘰𝘮𝘦 𝘱𝘦𝘳𝘴𝘱𝘦𝘤𝘵𝘪𝘷𝘦: 𝘵𝘩𝘦𝘳𝘦 𝘢𝘳𝘦 𝘻𝘦𝘳𝘰 𝘳𝘰𝘣𝘰𝘵𝘴 𝘪𝘯 𝘵𝘩𝘦 𝘸𝘪𝘭𝘥, 𝘢𝘵 𝘵𝘩𝘦 𝘮𝘰𝘮𝘦𝘯𝘵, 𝘵𝘩𝘢𝘵 𝘶𝘴𝘦 𝘢 𝘸𝘰𝘳𝘭𝘥 𝘮𝘰𝘥𝘦𝘭 𝘵𝘰 𝘸𝘰𝘳𝘬 𝘰𝘯 𝘵𝘢𝘴𝘬𝘴 𝘵𝘩𝘦𝘺 𝘸𝘦𝘳𝘦𝘯'𝘵 𝘴𝘱𝘦𝘤𝘪𝘧𝘪𝘤𝘢𝘭𝘭𝘺 𝘵𝘳𝘢𝘪𝘯𝘦𝘥 𝘰𝘯. 𝘛𝘩𝘪𝘴 𝘪𝘴 𝘤𝘶𝘵𝘵𝘪𝘯𝘨 𝘦𝘥𝘨𝘦 𝘳𝘦𝘴𝘦𝘢𝘳𝘤𝘩 ...”. - ricardobeat. 💡 If you’re passionate about physical world modeling, causal understanding, or building intelligent systems that interact with the real world—join us. More information, the paper, and a bunch of explanation videos to check here: 🔗 𝗦𝗶𝘁𝗲: https://ai.meta.com/vjepa 🔗 𝗕𝗹𝗼𝗴: https://lnkd.in/gwccegvT 🔗 𝗣𝗮𝗽𝗲𝗿: https://lnkd.in/gFGDi4xG 🔗 𝗠𝗲𝘁𝗮 𝗖𝗮𝗿𝗲𝗲𝗿𝘀: https://lnkd.in/g-MaBMc5 #AI #WorldModels #VJEPA #MetaAI #FAIR #EmbodiedAI #Robotics #CausalReasoning #LeCun #MachineLearning

161

10 Comments

The intersection of generative AI and the real world is finally becoming tangible. With Thumbtack’s new Alexa+ integration, homeowners can describe a problem in natural language and instantly connect with a pro who can fix it. This is what progress looks like: AI translating human intent into real-world action. We’re just scratching the surface of how natural language can reshape how people find, hire, and get things done. #Thumbtack #Alexa #GenerativeAI #Innovation #SmartHome #FutureOfWork

18

2 Comments

TL;DR - Tax code changes today are likely to lead to more hiring for R&D roles in the US, particularly for profitable high growth companies and big tech. Predictions: - Fewer layoffs in 2026 - Hiring for IC engineers recovers first - Trend towards off-shoring R&D slows, but does not stop or reverse. India and LatAm labor is still much cheaper and lots of great talent in those geos. https://lnkd.in/gyVqV6-j

20

As AI speeds up what’s technically possible, the real challenge is knowing what’s actually worth building. Great insights in this new article from our Chief Product Officer, Inbal Shani, on translating science into solutions.

1

Very interesting visualization by Andrej Karpathy about the US job market. Click the "Digital AI Exposure" button so see sectors that are most impacted by AI. Also, see the note about what that metric represents. https://karpathy.ai/jobs/

3

Put GPT-5 variants (including gpt-5-mini, gpt-5-nano) to the test against o4-mini and GPT-4.1. Task: Building my LinkedIn newsletter portfolio page. Same prompt, same requirements for all models - the UI quality difference is notable, much better with GPT-5. Fascinating to see how frontend capabilities have progressed across model generations. This is just a small sample - looking forward to testing GPT-5 on more complex workflows and sharing insights in my... newsletter. https://lnkd.in/gxhktdEF #GPT5 #AI

2

🎉 I just wrapped up the Evaluating AI Systems course by Hamel H. and Shreya Shankar - both practical and production-focused, covering the full AI application lifecycle from defining success metrics to operationalising evals in live systems. A ton of learnings, but the following takeaways stood out for me: 1️⃣ Start with the prompt: Even when prompts are clear and scoped, LLMs can still show biases, inconsistencies, or unexpected behavior with novel inputs. 2️⃣ Own the annotation process: Reviewing errors yourself is invaluable. In larger orgs, structured alignment sessions, using multiple perspectives, quantitative measures, and clear rubrics help ensure consistent, reliable evaluations. 3️⃣ Keep evals specific and monitored: Targeted evals paired with frequent monitoring are key to catching performance drift, especially degradation in judge or model behaviour. If you build AI/ML products as an engineer, PM, or leader, this course will give you the frameworks and playbook to make models measurable, reliable, and ship-ready. They're kicking off the next cohort on 6th of October, and you can use this link to access the course (with some discount): https://lnkd.in/eBxjfHWq

13

👀 Have you been watching the TV series 𝗧𝗵𝗲 𝗣𝗶𝘁𝘁? ✋ I have been binging (unfortunately, one of my worst vice). How about that AI notetaker that Dr. Al-Hashimi was trying to get the ER to 𝗔𝗗𝗢𝗣𝗧? Read more about it here: https://lnkd.in/dh272ygX This little bit is surprisingly so relevant to my work. Not because of the AI feature, but because I truly see how building great technology is only half the battle—adoption is everything. In healthcare, even the most innovative solutions face a tough reality once they leave the lab. It’s not just about performance metrics or clinical validation. It’s about fitting into real-world environments where time is limited, workflows are complex, and trust has to be earned. A few challenges that stand out:  • 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 & 𝘂𝘀𝗮𝗯𝗶𝗹𝗶𝘁𝘆 – If it’s not intuitive, it won’t be used consistently  • 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄 𝗶𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻 – Even small disruptions can slow down already stretched teams  • 𝗖𝗹𝗶𝗻𝗶𝗰𝗮𝗹 𝘁𝗿𝘂𝘀𝘁 – Adoption depends on confidence, not just capability What’s interesting is that success often comes down to the “non-technical” pieces—listening closely to users, iterating based on feedback, and designing with their day-to-day reality in mind. At Oncoscope, we understand these pain points. That's why: 👆we built our interface to be a 𝗷𝘂𝘀𝘁 𝗮 𝗳𝗲𝘄 𝘀𝗶𝗺𝗽𝗹𝗲 𝗰𝗹𝗶𝗰𝗸𝘀 from a living SLR evidence library to a customized PICOS SLR report 🌐 no training or downloading required, 𝗳𝘂𝗹𝗳𝗶𝗹𝗹 𝘆𝗼𝘂𝗿 𝗲𝘃𝗶𝗱𝗲𝗻𝗰𝗲 𝗻𝗲𝗲𝗱𝘀 𝗮𝗹𝗹 𝗮𝗰𝗰𝗲𝘀𝘀𝗲𝗱 𝗼𝗻𝗹𝗶𝗻𝗲 𝗳𝗿𝗼𝗺 𝗼𝗻𝗲 𝗽𝗹𝗮𝗰𝗲 🔗 𝘁𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝗰𝘆 of a clear SLR protocol and each individual publication traced back to the 𝗼𝗿𝗶𝗴𝗶𝗻𝗮𝗹 𝘀𝗼𝘂𝗿𝗰𝗲, and trust me, the numbers DO add up! Innovation in healthcare isn’t just about what we build. It’s about whether it actually gets used—and makes a difference when it matters most. Curious to hear from others in the space—what have you seen make or break adoption?

11

1 Comment