Agents, Metaprompting and Tech-powered Parenting with HubSpot CTO Dharmesh Shah

I recently sat down with Dharmesh Shah, co‑founder and CTO of HubSpot, on Canva’s Prompted podcast to unpack a set of ideas that feel both immediate and future-facing: AI agents, metaprompting, culture as a product, and even tech‑powered parenting. Dharmesh has always taken the unexpected path—starting companies on credit cards, writing one of the most-read culture codes in tech, and running a large engineering organization without a single direct report—and in our conversation he brought that same clear, pragmatic energy to the messy, exciting world of AI today.

I’m writing this as a distillation of that conversation: what Dharmesh believes, what he’s building, and what I think matters if you’re trying to bring AI into your organization or your life. I’ll share concrete examples (agent.ai, metaprompt.com, his “dad joke” experiments), practical playbooks (how I or Dharmesh would build an agent, what onboarding looks like), and the thoughts that keep him up at night—and the ones that keep him up with a smile.

Table of Contents

• Why I built a professional network for agents 🤖
• How I build agents and what they do 🛠️
• Can AI be creative? The simulation trick 🎨
• Agents as teammates or powerful tools? 👥
• Culture is a product — applying it to hybrid teams 🏗️
• Tech-powered parenting: how I introduced my son to AI 👨‍👦
• Metaprompting and practical AI tools I use 🔧
• Pricing, value and the economics of AI work 💸
• What worries me—risks and opportunities ⚖️
• The all‑star creative team I’d hire today 🌟
• Rapid‑fire takeaways and practical tips 🔎
• FAQ 🤔
• Conclusion: my bet on humans (for now) ✨

Why I built a professional network for agents 🤖

One of the first things Dharmesh told me was simple and brilliant: if we’re going to have AI agents on teams alongside humans, then all the institutions that help humans participate in professional life are going to need analogues for agents. How will you discover the right agent for a task? What will an agent’s résumé look like? How do you measure experience and trust?

That insight led to agent.ai: a professional network for agents. It started as a thought experiment—there should be a place where people can find agents, share them, and evaluate them—and has evolved into a platform where folks are already building and publishing agents. Dharmesh told me they launched in beta around 50,000 users and that the platform now has roughly 2 million users and 2,000 publicly shared agents. Those numbers show momentum, but more importantly they point to real behavior: people are experimenting by crafting agents, wiring APIs, and sharing the results.

Why does this matter? Because agents aren't just single-purpose chatbots anymore. They're compositions: web scraping + proprietary APIs + LLMs + deterministic workflow logic. They can act as a researcher for you, a prospecting assistant, or a customizable report generator that you call into whenever you want fresh intelligence on a company or market. That's the world Dharmesh expects and one he’s building for.

How I build agents and what they do 🛠️

When I dug into the mechanics with Dharmesh, he described a spectrum of approaches. At one end there are low-code, deterministic flows: explicit multi-step sequences with inputs, decision nodes, and calls to LLMs and other APIs. That approach is great for predictable tasks that require governance or repeatable outcomes. At the other end are full-featured, off-platform agents—built with “agentic” coding frameworks of your choice—that plug into the agent.ai ecosystem.

Here are the core building blocks he highlighted:

• Deterministic flows: Low-code building blocks that define the steps an agent takes, the inputs it needs, and the outputs it should produce. Useful when you need predictable behavior and auditing.
• LLM calls: The language model is often called as a function in the workflow to do summarization, reasoning, or content generation.
• Data connectors: Social media, company data, proprietary APIs—these provide the external inputs the agent uses to make decisions.
• Custom UIs: Minimalist UIs for users who don’t need to see the plumbing, and full custom interfaces when the task demands higher interaction fidelity.
• Agent-to-agent wiring: Agents can bring other agents into a flow when the task benefits from cross-domain knowledge.

Dharmesh walked me through one of his favorite agents: the company research agent. You type a company name or domain and the agent gathers signals from the web and proprietary sources, synthesizes the findings, and then customizes the report based on the questions you care about. If you’re in a specific vertical, you can tell the agent the kinds of metrics, compliance items, or business models you care about and it will tailor the output. That’s infinitely more useful than a generic summary because it maps to human intent.

He also confessed, with a laugh, to a very human project: a dad‑joke agent. Despite every advance in prompting and reasoning models, his experiments showed that LLMs still struggle to invent an original, decent dad joke. They can explain why a joke works, but they’re not consistently great at producing or scoring one quantitatively. That anecdote tells you two things: first, there are still gaps in AI creativity for very specific flavors of humor; second, good agent design includes knowing where a model is strong and where humans still have the edge.

Can AI be creative? The simulation trick 🎨

We spent a long time on creativity because this is where people get excited and anxious in equal measure. Dharmesh reframed creativity in a way I find useful: creativity is often about connecting primitives—ideas, elements, concepts—in novel combinations. AI, conditioned on enormous amounts of data, is naturally a connection machine: it creates proximity between ideas humans might never have considered linking.

But Dharmesh didn’t stop at that definition. He shared a practical technique his son uses: worldbuilding as simulation. His son—an aspiring fantasy author—builds a long, detailed prompt that defines the world, characters, laws of magic, and political structure. Then he uses the model to pressure-test the logic: what happens if character X makes this choice? What emergent conflicts appear? That iterative simulation helps surface narrative threads that the author might have missed.

I liked that example because it highlights two things AI does well:

1. It can run through permutations and surface internal contradictions or opportunities quickly.
2. It acts as a critic and brainstorming partner that scales—giving hundreds of scenario checks in the time it would take a human to write one.

Dharmesh also stressed that he uses AI as a brainstorming tool and critic rather than as a final author. He’ll use models to improve transitions, suggest options, or propose ways to connect disparate product features. Often the output is a mix of usable and unusable ideas, and the human’s job is to find the “gold”—to act as an editor and curator.

On the humor front, the dad-joke exercise revealed a subtlety: models can analytically explain why a joke works but still falter at producing or measuring one. Dharmesh tried rating a curated set of jokes and the model’s scores didn’t align with expectations. That suggests that subjective metrics—like comedic timing or cultural resonance—are still hard to quantify automatically. One possible workaround he floated: ask the model to simulate an audience of 1,000 people with specific demographics and estimate how many would laugh. That kind of proxy might be more useful than asking the model to score humor directly.

Agents as teammates or powerful tools? 👥

One of the big questions everyone asks is whether agents will become teammates—trusted, autonomous collaborators—or remain powerful, programmable tools. Dharmesh’s answer was clear and thoughtful: in raw capability, many agents already have the IQ to do tasks we traditionally think of as human. But the transition from tool to teammate isn’t purely technical; it’s organizational.

He gave a metaphor I loved: think of an AI agent as an extremely adept intern. It can be brilliant at knowledge and reasoning, but when it joins a company it doesn’t know the business, the policies, the culture, or the teammates. You wouldn’t expect a brilliant intern to be fully productive without onboarding, feedback, and context. The same applies to agents.

So the challenge isn’t just to build smarter models—it’s to design the systems that let agents absorb context, measure and improve performance, and integrate as members of a hybrid team. Dharmesh highlighted a few practical needs:

• Onboarding: Agents need training on company-specific documents, norms, and systems so they aren’t “unleashed” without guardrails.
• Feedback loops: Agents should get measurable signals about how well they’re performing—akin to manager feedback—and they should be able to improve with iteration.
• Discoverability: Agents should be able to find and call other agents when cross-domain help is needed; agent-to-agent protocols will be crucial here.
• Embedding: Agents must live where work happens—email, Slack, CRMs—so they aren’t siloed behind separate tools or dashboards.

He also believed that the major unlocking moment will come when agents have knowledge of one another and can pull in complementary expertise. Today’s agents tend to stay “in their lane”; tomorrow’s networked agents could surprise us by answering questions they weren’t explicitly programmed to handle.

Finally, even when agents become “teammates,” Dharmesh stressed that humans will still be the glue. He’s optimistic about agents handling many tasks, but he bets on humans—humans who know how to orchestrate, coach, and apply taste. In his words: “My money is still on the humans. Humans powered by AI.”

Culture is a product — applying it to hybrid teams 🏗️

Dharmesh’s view on culture is one of those simple statements you remember: “Culture is a product.” He means it literally: just as you’d build a product for customers, you build culture for your team. That mindset changes the behavior of leaders. If culture is a product, then you measure usage, gather feedback, iterate, and rethink who the customers are over time.

Two corollaries followed:

1. Culture evolves. The cultural product you need at 50 people is not the same product you need at 5,000 people.
2. Culture requires metrics. If a product should have customer feedback, culture should have it too—so Dharmesh recommends the same discipline we apply to product: ongoing measurement like employee NPS and listening systems.

In the era of hybrid teams—humans plus agents—he argues there are now two sets of customers for the culture product: people and agents. That raises complex questions. Should you design culture that optimizes for both? Does an agent get a vote on company values? His short answers: be mindful of agents as constituents, but humans remain the primary customers of culture. Agents may have a voice—through telemetry, logs, or usage patterns—but they don’t get a vote in the human-centric sense.

Practically, this means starting to instrument your culture for agent readiness. You might:

• Audit processes to ensure they’re machine-interpretable where it makes sense.
• Design onboarding that brings agents up to speed on the organization’s policies and expectations.
• Collect agent-centric metrics: how often a support agent needs human escalation, whether an agent’s suggestions lead to actions, or how an agent’s performance changes as it receives feedback.

Thinking of culture as a product reframes the conversation from “Do we like this old startup vibe?” to “Is the team getting what it needs to do its best work?” That shift is crucial as organizations adopt hybrid workflows.

Tech-powered parenting: how I introduced my son to AI 👨‍👦

Dharmesh’s approach to parenting in an AI world is refreshingly pragmatic. He exposed his son to technology early—his son had access to models like GPT‑2 well before ChatGPT became ubiquitous—and has encouraged responsible use rather than sheltering him from the tech.

Two big ideas guide his parenting philosophy:

1. Blank-slate optimism: Kids assume technology works. They don’t have a pre‑internet world to compare it to. That often produces wonderfully creative prompts and approaches because they’re not constrained by an “old way” of thinking.
2. Guidance, not prohibition: Rather than banning tools, Dharmesh shows his son how he uses AI, what he expects from it, and where he draws lines. The goal is to develop judgment and responsibility.

One delightful detail: his son loves treating AI as an “arbitrary, objective third party.” When a kid thinks a parent is biased, a neutral model can provide a different perspective—and that can catalyze a conversation that’s better than an argument. Dharmesh uses AI to get parenting ideas, troubleshoot uncommon problems (because AI has seen many parenting scenarios), and to spark dialogue. He doesn’t see AI as a replacement for parental judgment—just as a tool that expands what’s possible in the conversation.

We also touched on creative use cases: his son uses long, 2,000-word prompts to build fantasy worlds, which the model iterates on with him. That is skill-building. He’s getting better at translating the ideas in his head into a tangible output, using a tool that lowers the technical or craft barrier to writing and design. In Dharmesh’s view, that’s the upside of lowering the bar: more people can create, and more diverse creative voices will emerge.

Metaprompting and practical AI tools I use 🔧

One of Dharmesh’s contributions that I think deserves immediate attention is the idea of metaprompting—and the practical tool metaprompt.com. Prompt engineering has become a craft because good prompts tend to yield better outputs. Metaprompting is a step beyond: use AI to make your prompt better.

Here’s how Dharmesh’s metaprompt approach works in practice:

1. You paste the prompt you use frequently into the tool.
2. The system asks clarifying questions: “What do you want this style to be? What constraints? Who’s the audience?”
3. Once you answer, it rewrites your prompt into a cleaner, better-structured, more likely-to-succeed prompt you can reuse.

That’s a one-time investment with outsized returns. If you ship many prompts—weekly newsletters, sales outreach, research queries—tuning the prompt once can save hours of iteration and lift the quality of every output.

Dharmesh also talked about the other AI tools he uses:

• Willow Voice: A desktop app that provides voice input across apps and formats text with app-awareness (e.g., switching tone for email vs. notes).
• Perplexity: A search-based AI tool Dharmesh uses for targeted informational prompts (“Tell me about person X or Y”).
• AI for editing: He typically uses models for editing and copy polishing rather than full writing—particularly for headline ideas or punchy subject lines.
• Visual AI: For people like me who aren’t designers, AI tools make it easy to manifest visual ideas and lower the bar for creating marketing assets, presentations, and blog visuals.

These are pragmatic uses: they don’t try to replace the human spark, but they increase throughput and help shape the voice and look of output more efficiently.

Pricing, value and the economics of AI work 💸

One of the trickiest things for businesses is figuring out how to charge for AI-enabled work. Dharmesh walked me through the different models and where they make sense.

Outcome-based pricing is seductive: charge for solved tickets, conversions, or saved hours. It works well in places where outcomes are objective and measurable. Customer support is a classic example: did the ticket get resolved? Was CSAT maintained or improved? There’s a clear analogy to the old metrics for human-run processes, and the economics are placeable: you know how much a human ticket costs and can price the automated version accordingly.

But outcome-based pricing breaks down for many creative, subjective, or nonlinear tasks. Consider logo design. What does it mean to pay for the outcome? You can create a test to measure audience preference, but the economic value of a single “brilliant” logo can be orders of magnitude different than a run-of-the-mill one. A lot of human value in creative work is about taste, intuition, and cultural resonance—things not easily reduced to the average of many outputs.

Dharmesh’s bottom line on monetization: outcome pricing will work for certain fungible, measurable tasks. For others, tools that increase human productivity and creativity will continue to be priced like tools (subscription, per-seat, consumption). There’s still a huge role for humans in adding taste and nonlinear value: my money is still on humans—especially humans who are powered by AI.

What worries me—risks and opportunities ⚖️

So what keeps Dharmesh up at night? He split his concerns into a dark side and a lighter, optimistic side.

On the darker side: he’s less worried about some hypothetical doomsday AI takeover and more worried about how bad actors will weaponize highly believable, multimodal generative content. Imagine voice, video, images, and text that are all indistinguishable from reality. Malicious people with motives—and now with these power tools—can do more harm, faster. The solution here, for Dharmesh, is education: we need citizens to understand what modern AI can and cannot do and to build systems that verify provenance, check authenticity, and raise general awareness.

On the lighter side: he sees history repeating. He felt the same excitement and frustration during the early internet days—millions could benefit but many didn’t because the tools and education weren’t there. He thinks AI is in that phase: powerful, but underutilized. Millions of people use chat applications but only scratch the surface of what these systems can do. The opportunity is to democratize access and make powerful AI usable by non-experts. That’s exactly the mission behind projects like Simple and some of HubSpot's AI efforts: build tools that make it obvious and easy to get value from AI.

The all‑star creative team I’d hire today 🌟

As a fun aside—because I always love asking this—I asked Dharmesh to pick his four-person creative dream team. He chose people who are alive because he likes constraints and hopes he might actually work with them. His picks were instructive because they mix artistry, design, product thinking, and technical explanation:

• Jacob Collier — a multi-instrument musician and creative virtuoso who Dharmesh called a “product guy” in the sense of serving an audience while bringing deep craft. Jacob’s ability to orchestrate complex musical ideas made him an interesting pick for creative leadership.
• Jack Butcher — the visual thinker behind Visualize Value, known for reductive, elegant visual abstractions. Dharmesh admires Jack’s capacity to strip complex ideas into simple, memorable visuals—something invaluable in product and design.
• Jason Fried — of Basecamp (37Signals), known for clarity in product and writing. Dharmesh values Jason’s ability to distill ideas and make them operational: he’s a product and organizational thinker.
• Andrej Karpathy — a technical leader who can translate complex deep learning concepts into language people can understand. Dharmesh admires this skill because it bridges the chasm between research and real-world adoption.

Why this group matters: it’s the combination of craft, clarity, product focus, and technical distillation that makes modern teams sing. Dharmesh’s picks are a reminder that the future of creativity mixes multiple disciplines—and that the right team can turn complexity into something consumable and useful.

Rapid‑fire takeaways and practical tips 🔎

If you want to take action today—whether you’re a founder, manager, designer, or parent—here are the distilled playbooks and heuristics Dharmesh shared, reordered as practical advice you can apply immediately:

• Start with a clear problem: When building an agent, define the job-to-be-done. Is it research, prospecting, ticket resolution? Clarity beats shiny tech.
• Use low-code for governance: If predictability and auditability matter, start with deterministic flows that call LLMs as functions rather than giving models free rein.
• Instrument onboarding: Treat agents like interns: give them onboarding materials, policies, and measurable feedback so they improve instead of flailing.
• Metaprompt once, reuse forever: Invest time in making your prompt excellent. Use meta-prompting to transform brittle prompts into robust templates that scale.
• Embed agents where work happens: Agents should live in Slack, email, CRMs—don’t make people jump to a separate dashboard to get value.
• Use AI as a critic and connector: AI is fantastic at generating permutations and connecting disparate ideas; use it for simulation and brainstorming, not as a final author.
• Measure outcomes where possible: For fungible tasks (tickets, data entry), outcome-based pricing works. For subjective creative tasks, measure carefully and be skeptical of pure outcome pricing.
• Teach kids responsible use: Expose them to the tech with guidance rather than prohibition. Encourage creativity and skepticism at the same time.
• Remember human taste: Machines don’t yet substitute for human judgment in high‑taste domains (branding, culture, creative leadership).

FAQ 🤔

Q: What is an AI agent, and how is it different from a chatbot?

A: In practice, an AI agent is a composed system: LLM capabilities wrapped around deterministic workflows, data connectors, APIs, and sometimes custom UIs. A chatbot is primarily an interface for conversational interaction. Agents are built to do work autonomously or semi-autonomously: fetch information from disparate sources, execute multi-step workflows, call other services, and—crucially—be discoverable and reusable. Think of an agent as a role on your team rather than just a chat window.

Q: Will agents replace jobs?

A: Agents will replace tasks more than jobs. Repetitive, predictable tasks are most at risk. But humans bring taste, context, and non-linear judgment. The real opportunity is to shift people away from tedious work toward higher-level strategy, creativity, and orchestration. The leaders who embrace hybrid teams—humans plus agents—will get disproportionate leverage.

Q: Can AI really be creative, or does it only remix?

A: It depends how you define creativity. If creativity is novel recombination of primitives, AI is already incredibly good. If creativity requires lived human experience that can be mined for original metaphors, then AI is not yet a full substitute. The pragmatic path is to use AI for ideation, simulation, and iteration, and keep humans in the loop as editors and curators.

Q: What’s metaprompting and why should I care?

A: Metaprompting is the practice of using AI to improve your prompts. Instead of manually tuning prompts forever, you use a model to ask clarifying questions and generate a robust prompt that fits your intent. That single investment can dramatically improve the quality of future outputs and save time across repeated tasks.

Q: Are agents trustworthy? How do you measure them?

A: Trustworthiness is earned like with humans: onboarding, clear objectives, feedback loops, and measurable KPIs. For customer support agents, you can measure resolution rate, CSAT, and escalation frequency. For research agents, measure accuracy, relevance, and actionability. Build instrumentation so you can audit decisions and have human oversight for risky domains.

Q: Should I give AI a vote in company decisions or culture?

A: My short answer: no. Humans should retain decision-making authority when it comes to values and culture. Agents can provide telemetry, feedback, and data-driven suggestions, but culture is fundamentally a human product aimed at human customers. Treat agents as constituents whose experiences you should measure, but not as voters that shape human values.

Q: How do I start introducing agents into our team?

A: Start small. Pick a well-defined, measurable use case (e.g., lead enrichment, research briefs, or internal knowledge triage). Build a deterministic flow that calls an LLM for the uncertain step. Instrument outcomes. Create a playbook for human oversight. Iterate based on real usage, and scale once you’ve nailed governance and output quality.

Q: What’s one simple experiment I can run this week?

A: Try metaprompting for a repeated prompt you use. Put your current prompt into a tool like metaprompt.com (or ask an LLM to improve it), answer the clarifying questions, and then use the revised prompt for a week. Measure whether the outputs are more useful or require less editing. That single experiment often unlocks immediate productivity gains.

Q: What about the ethical risks of AI models generating fake media?

A: The real risk comes from malicious use of highly realistic media. The obvious defenses are verification systems, provenance mechanisms, and public education. We should invest in tools that can verify whether content is synthetic and in public campaigns that teach people critical consumption patterns. Regulation and industry standards will also have roles to play.

Conclusion: my bet on humans (for now) ✨

There’s a healthy tension in the conversation we had: the technical competence of models is rapidly accelerating, and yet the human element—taste, judgment, moral responsibility, context—still matters enormously. Dharmesh’s thesis is beautifully pragmatic: build infrastructure (agent networks, onboarding, feedback), democratize AI tools so more people benefit (a repeat of the internet moment), and remember that for many high‑value problems the human touch remains essential.

I walked away from our talk energized by the practical things you can do today—metaprompting, instrumenting culture, treating agents as interns that need onboarding—and grounded by a reminder that the best outcomes will probably come from hybrid systems: humans who know how to harness AI, and agents who are built to play well inside organizational contexts.

If you’re building with AI, focus on three things I heard again and again in our conversation:

1. Make the problem explicit before you start wiring models into systems.
2. Invest in the non-technical pieces—onboarding, feedback, integration into daily tools—because those are what turn a clever script into a reliable teammate.
3. Keep humans in charge of taste and values. Use AI to extend judgment, not replace it.

And finally: if there’s one charming experiment you should try, it’s this—ask your model to write a dad joke, then try to tell it to your team. The model might explain why a joke works, but judging from Dharmesh’s ongoing crusade, it still might not pass the “is this actually funny?” test. It’s a small, delightful reminder that while AI is extraordinary, the human sense of humor—and the messy, unpredictable qualities that make us human—still matter.

“Culture is a product.” — Dharmesh Shah

If you want to dig deeper, Dharmesh has launched tools like metaprompt.com and agent.ai, and if you’re curious about the organizational side, you can’t go wrong internalizing the idea that culture needs measurement and iteration. I’ll close by saying: I’m optimistic. The next year will be full of experiments that move us from curiosity to competence. If you’re building, start small, instrument rigorously, and keep your human instincts sharp—because those instincts are still the competitive advantage.

---