Dharmesh Shah's Recent LinkedIn Posts

This has a great "wow-to-work" ratio. You get a lot of "wow -- that's useful" with very little work. If you work in sales/growth, you should check it out.

Current Status: Dealing with the emotional roller coaster that is the Anthropic API right now (they're having an owie). On the one hand, I'm in a state of flow and trying to finish up some work for a demo video I want to record a screencast for. On the other hand, it's draining to keep retrying things 5 times in order to get it to work. Just creeping along. My best wishes to the Anthropic team. Never fun to be going through system issues -- especially on a Friday afternoon.

I've had this rule in my professional life for 30+ years. The "space" you give yourself when it comes to meetings/calls/discussions should be proportional to the potential and stakes involved. This is especially true when you have much less experience/savvy you're bringing to the table and they have much less value they're likely to get out of it. Example: In the early years of HubSpot, when Brian Halligan and I flew out to the west coast for fund-raising trips, we would *never* fly the day of. We would fly out the night before. We would hang out at the hotel and have a "founders meeting" late into the night to chat about the meetings and get into the right headspace. And we'd leave ourselves space between meetings to reflect, adjust and prepare for the next one.

My current sense of things when using AI to build product: GPT 5.4 is the best PM. Best at reasoning/thinking and long-range execution. Lovable is best UX designer for creating visually attractive prototypes. Opus 4.6 is the best front-end/UX engineer. Combination of taste and technical skills. GPT 5.4 is the best back-end engineer and architect. It thinks things through more deeply and is more precise in execution. Let me know if your experience is different -- or if I missed something.

Did a recent podcast appearance with Rowan Cheung of "The Rundown" fame (full YouTube video in comments). Some of the ideas discussed: * The best way to be rank in Google is to be rank-worthy. The best way to get cited in ChatGPT/Perplexity/etc. is to be citation-worthy. * Treat agents as if they were teammates. You wouldn’t hire an intern... and just hand them a computer and say, 'Here’s what we want done, go do it.' You'd do some onboarding. You'd train them. You'd give them regular feedback. Same for agents. * Don't try to coerce a reasoning model into figuring out steps you already know need to happen. If you have already reasoned through the steps... build a workflow agent. Hope you enjoy it. Let me know if you have any questions/comments.

Agent In The Arena It is not the commentator who counts; not the analyst explaining what the agent should have done; not the architect describing how the system might have been designed better. The impact comes from the agent in the arena — the one executing tasks and achieving outcomes. The one operating in the wild where context is incomplete, tools misbehave, and skills come up short. Not the pundits, prognosticators and perpetual pessimists. The impact comes from agents and the builders who build them. The agent in the arena is the one whose runs are marked by retries and rate limits; the one that attempts then reasons, reroutes and retries. It will hallucinate sometimes. It will fall and trip sometimes. It will outright fail sometimes. But the builder learns, adjusts, and tries again — iteration by iteration. The arena is not the meeting room. The arena is not the demo. The arena is production. At its best, the agent amplifies human intent. It pairs human judgment with machine execution so that together they move work from possible to practical. And at its worst, if it fails, it fails while pushing learning forward, task by task, query by query, attempt by attempt. Because the future will not be forged by the fearful. It will be wrought by agents doing the work. The last era of AI was about models that could understand the world. The next era belongs to agents that can change it. h/t "Man In The Arena" by Theodore Roosevelt

Woo hoo! Thrilled to share a sneak peek of HubCode -- The Vibe Coding Tool for HubSpot. I talked about this a few weeks ago (which feels like forever). Been cranking on it and am ready to show you the actual app now. Link to video in the comments. Or, just type video.hubcode .com in your browser (it takes you to YouTube). Some notes: 1) This requires no terminal -- runs completely in the cloud 2) You can vibe code both app cards and full apps that live inside the HubSpot web UI. 3) Generated apps can access data within HubSpot (of course). 4) SUPER EXCITING: Apps can reuse any of the public agents in agent .ai (there are 2,000+ of them). This makes your apps super-powerful because you can access all sorts of data and services. And, you can build your own agents and access them too. This gives you access to all the major frontier models, a bunch of really useful data sources and a ton of other capability. Go to agent .ai to see what's available. The demo shows pulling YouTube videos related to a company right into the app card. 5) You can paste in screenshots or other visuals as part of your vibe coding process (if you haven't done this before, it will feel magical). The HubCode app runs on the agent .ai platform and is in private beta (because this one's going to take some testing and iteration give the power and flexibility). To apply, just go to hubcode .com. This is an early "proof of concept", but will be iterating on it maniacally as I get user feedback. Let me know what you think. All thoughts, ideas, feedback and wishlist items appreciated. FEEDBACK IS A GIFT. Thank you! Thank you for your support.

Wrote this 18 years ago (2008). I think most of the advice aged pretty well. My favorite tip: Allocate Most Time To Customer Value:  Act as if someone is paying you $1,000/hour for every hour you spend making life measurably better for your customers -- and $10/hour for everything else.  In the long-run, the ratio will be about right.

From a developer at HubSpot: --- Hi Dharmesh, I've got a friend that's got a year left in their PhD program at MIT and has some spare cycles. They've got off-the-charts Python skills, and are really good at Java and Typescript too. And SQL. And they know how to leverage LLMs. We've got a bunch of work on our product roadmap that I think they could crank through, no problem. They don't care about money, but I'd want to pay them something. Maybe $25/hour. OK to do this and get them on board? I'll take responsibility for managing them and keeping them productive. What do you think? --- I'll bet you can guess my likely response. Sounds awesome! Let's get them started today and take as many hours as they have available. No limit. If you think they're that good and you can make them productive, I'd be thrilled if they worked 12 hours a day, 7 days a week. Whatever time they have available, we'll take those hours. --- Chances are, you'd do the same if you were me. Now, what if I told you this was a hypothetical email and that the associate the developer was looking to hire was named Claude? Or Codex? Or Cursor? Would that change the answer? It wouldn't, right? Because it just makes sense. If I can help a developer be more productive by letting them leverage an AI associate (i.e. an agentic coding tool), why wouldn't I do that? We have the opportunity to give every one of our people an associate/intern for less than $25/hour. And they are *awesome*. Some questions for your consideration: 1) Is the developer that can manage such an agentic coding tool more valuable? 2) Would you put a limit on how much the developer could use AI as long as they were continuing to be productively profitable? 3) If more developers on your team came to you with such an offer, would you let them bring on associates too? As we speak, this debate, discussion and deliberation is happening in companies everywhere. My quick take: Spending on AI tokens to amplify human talent is the deal of the decade.

HubSpot's Breeze AI Assistant (accessible in-product) has come a long way since it first started as ChatSpot 3 years ago, way back in 2023. Minutes ago, I needed a custom property on the Company object in my personal HubSpot portal (I'm doing some tinkering with Custom Workflow Actions). So, I asked Breeze to do it with the prompt "Create a custom property on company record named agentai_test). It told me what it was going to do, and asked me to confirm. I did and BOOM -- I now have the requested custom property! I know that seems like a minor thing, but it brought me joy.

For me, the increase to 1M token context window when agentic coding is not about the model being able to deal with more code. It's about me being able to crank through a discrete task without having context anxiety. Before, I'd often get really close to compaction as I was working through a task and would worry that something important would be lost. Now, I almost never hit that limit for any given task that I'm working through.

We live in magical times. I can now fly coast-to-coast and still have power to spare on my laptop (MacBook Pro). I have Internet access most of the time. I can run a local LLM with access to a decent percent of public human knowledge and chat with it in natural language.

I think about AI agents a whole lot. In order to have them fulfill their full potential, they're going to need the infrastructure and tools to let them do what they need to do. One of those is having an email account. So, I'm thrilled to be an investor in AgentMail. I'm also a customer (my first use case was to create an email account for a couple of my OpenClaw instances).

Have I mentioned before that I really love the idea of simple AI agents working together as a team to help people with a particular task? Here's the latest from Agent AI -- for market research. Are there other tools to do this? Sure. But not all are supported by an amazing team that are here to help you grow better with AI.

Excited to be supporting MIT delta v this year as we kick start a new chapter for the accelerator. Looking forward to spending time directly with the cohort as part of its Inaugural Partner network. For those unfamiliar, delta v is MIT’s flagship startup accelerator for student founders, and the 2026 program has been redesigned to focus on a smaller group of highly committed teams building audacious companies. Selected teams receive: • $75K in equity-free funding  • Over $200K in perks and partnerships  • Direct access to founders, operators, and investors through the new Partner model  • A full-time, in-person summer building at the Martin Trust Center in Boston The goal is simple: help teams spend the summer building something real that customers want and will pay for. Link to more details and the application in the comments.

BREAKING NEWS: Meta buys Dreamer. Boy that was fast! Meta just made a quiet move with loud implications. They’ve effectively acqui-hired the team behind Dreamer — a startup focused on helping people build autonomous agents — including co-founder Hugo Barra (Google/Meta) and David Singleton (former CTO of Stripe). Two details matter more than the headline: This is about people, not product. Dreamer’s tech isn’t coming along for the ride. The destination is agents, not chat. Put those together and the pattern becomes pretty clear. Meta isn’t just building better chatbots. They’re assembling a roster to build the agent layer. Here's my take: If chat was the “UI” of the first wave of AI, agents are the “UX” of the next. From “ask me a question” → “get this done.” That shift is bigger than it sounds. Because once agents become the primary interface: Apps become tools for agents. Workflows become programmable. Attention shifts from screens to outcomes. In other words, the battle isn’t for better answers. It’s for better execution. Billions of users won’t just interact with AI…they’ll delegate to agents. That's very much the world we're building towards at HubSpot with our Agentic Customer Platform. It'll be the "operating system" for GTM agents (both ours and others) all working off a common set of shared primitives and rich context. We are at the start of the agentic transformation.

The best way I learn is to to just DO THE THING and interact directly with customers. I love being able to own an early project and go into "Full Stack Founder" mode where I do everything from marketing, development -- and support. Currently working through feedback from beta users of HubCode (the agentic coding tool for HubSpot). You can sign-up for private beta at HubCode .com. Lesson for the day: Although I had a way to "restore" the user prompt to a prior version (allowing better experimentation), that's not enough. 1) Because the system is non-deterministic, the same prompt a user used before will not generate the exact same app code (after restoring it). This was irritating/confusing users. 2) So, a better answer is to implement a way to 'snapshot' a project that is deployed so the user can go back to a prior version/build of the project. 3) Might consider a variation of the /rewind feature in Claude Code whereby the user can go back on a prompt + project files. For those of you that have been using the beta, please keep the feedback coming. If you're still on the waitlist (there are hundreds of you -- sorry), best way to move up the line is to go through steps in the tutorial on HubCode .com. Further you are in terms of getting initial setup done (like connecting your HubSpot portal), the more quickly you will get approved. Thanks for your support. #hubcode

This is just the beginning. The HubSpot team is on 🔥!

Makes me happy to see teams of easy-to-use AI agents being launched on Agent.ai. AI agents don't have to be complicated. Think of them as smart software that can complete tasks that software really couldn't do before.

I'm "knows that long before markdown there was markup and that it's the M in HTML" years old.

The HubSpot team was at an offsite this week. Best 2 days of my year so far. So much to talk about and discuss and I'm biased, but the team is amazingly clueful. On the AI front, here's something that I think everyone that is developing with AI (and who isn't?) should consider: It's not just that the models are getting smarter (which they are), it's that the capabilities *around* the model that the frontier model companies are shipping are getting more capable. Tool calling, code execution sandboxes, skills, clawification etc. To make the most use of what's available in AI, you can't just focus on the core model. You need to understand what's available *around* it to really make the most use of it. Otherwise, you end up doing a bunch of work that doesn't really need to be done. This creates a new challenge: Way back when, when using the model APIs was just about text in / text out (completions API), switching across models was trivial. The way you invoked the model was the same regardless of which model it was. But now as we get new "features" around the model, we're going to see some variance in terms of *how* these features are used, and what capabilities they each have. An incidental byproduct of this is that developers/builders are going to have some incentives to standardize on *one* model company (OpenAI vs. Anthropic) to best leverage the capabilities of that company. Reminds me of cloud computing back in the day. Some companies would use just the very basic primitives from AWS (compute, storage), so they could maintain optionality, while others would use higher-level abstraction offerings. There's no "right" answer. It's just important to understand the tradeoffs.

Hello, San Francisco! Here a day later than planned but nice to be back. Sadly, a super-tight schedule this trip so won't be doing my usual founder meet-ups and dinner and such. But, I'll be back.

I've been working and building in the CRM industry for 30+ years (so have far exceeded the needed 10,000 hours). The last big transformation we saw in CRM was over 25 years ago with Salesforce's launch of what became the Cloud CRM. Eventually, every (successful) CRM was a cloud CRM. The next big transformation is happening now with the advent of AI and agents. But it's not about being "AI first", it's about being CONTEXT FIRST. Context isn't a feature. It's the whole game. Modern AI models are sensationally smart. But success is not just about high IQ, it's also about having high CQ (Context Quotient). The smartest person in the room is useless if they just walked in. An AI that knows your Q2 pipeline is full, your best rep is on parental leave, and your biggest account just hired a new decision-maker responds very differently than one that doesn't. In most companies, context lives in databases, docs, message threads… and people's heads. It's scattered and fragmented. That's a problem because without shared context, AI is just a very smart intern on their first day at work. AI agents are awesome – but only if they're context aware. So, I'm thrilled to finally share what HubSpot has been working towards. It's been 20 years in the making: The Agentic Customer Platform A customer platform built for both humans *and* AI agents. One that is context-first. A platform that combines the world's smartest AI models with the deepest context to deliver the most effective agents to drive your growth. Agents are the future of software and agentic is the future of customer platforms. Eventually, every (successful) customer platform will be an agentic customer platform -- and every successful GTM agent will need to integrate with an agentic customer platform. Yes, I know I'm biased, but that doesn't necessarily mean I'm wrong. :) You can read more details about our vision in a post today by Yamini Rangan (HubSpot's CEO). You can get to it by visiting: acp .net (yes, I like short links…and I don't know why). I'll be digging into the details of what this means from a product/technology perspective and how it actually works over the coming weeks and months. I love it when the dots start to connect. If you have any questions, comments or feedback, would love to hear them.

Anyone else seeing a dramatic increase in their LinkedIn follower count? Over the past 48 hours or so, I've gained 10,000+ followers. This is 10X+ higher than my usual pace of growth. Feels like it could be a technical issue, because none of my recent posts have done *that* well. UPDATE (03/16/2026): Looks like it may have been a technical glitch in the follower count. Now showing back at "normal" levels. Oh well...back to building.

Currently pondering: Should there be a "File System Protocol" whereby systems make their data/context available via a standard protocol that AI apps/assistants/agents can navigate and use? Kind of like MCP (Model Context Protocol) but different and even simpler. Last week I wrote about how coding agents were really, really good at using CLI tools -- and chaining them together. This week, I'm thinking about file systems and how coding agents are *also* very good at effectively using them. They're good at understanding the hierarchical structure, navigating the tree, doing searches, etc. The nice thing about file systems is that they're based on a relatively simple set of primitives but can represent a relatively rich, auto-discoverable data store. And they're familiar and friendly to humans too. Decades ago, NFS (Network File System) existed, which was all about making a file system accessible remotely. I'm not an NFS geek, but wondering if a modern protocol based on the needs of today's AI agents could be useful. We've had things like GraphQL available too, which lets you represent a traversable graph for systems to access data/context as well. Not sure which is better. Graphs are richer (and I happen to love them), but file systems are *simple*, and already exist in operating systems that we use on computers (and things like OpenClaw use them for all sorts of things -- including memory). No concrete thoughts. Just some midnight hour #pondering. (I'll do just about anything to avoid recording the video I was planning on recording today).