Mar 28, 2025 · 1h 42m · latent-space
The Agent Network — Dharmesh Shah, Agent.ai + CTO of HubSpot
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
HubSpot CTO and Agent.ai founder Dharmesh Shah joins Swyx and Alessio Fanelli to discuss the technical architecture of multi-agent networks, open protocols like MCP, pragmatic software engineering, and the future role of human developers. He shares deep strategic insights on knowledge graphs, asynchronous productivity, agent memory, and emerging business models in artificial intelligence.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 7.4% of the talking time here. How this is scored →
The hosts as informed peer 5.6 Guest teaching 4.9 Guest disagreement 1.5 The hosts pushing back 2.1
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 50:14 Dharmesh rejects replacing Python with dedicated AI languages
Dharmesh explicitly takes issue with Bret Taylor's thesis that AI requires a brand new programming language, firmly asserting Python remains the optimal bridge between human and machine comprehension.
Hardest push from the hosts ▶ 18:31 Swyx challenges the practical viability of testing graph databasesSwyx directly pushes back on Dharmesh's optimism regarding graph data stores, arguing developers cannot realistically benchmark multiple experimental options due to resource constraints and eval scaling limitations.
Biggest teaching moment ▶ 1:03:00 Dharmesh analyzes why Results-as-a-Service fails outside narrow verticalsDharmesh educates the hosts on the economic realities of pricing models, explaining why outcome-based pricing thrives in standardized customer support but collapses in subjective or highly variable workflows like design.
The host holds their own ▶ 1:34:38 Swyx details insider shifts in autoregressive multimodal architecturesSwyx showcases superior technical industry context by detailing how Google Gemini abandoned diffusion in favor of autoregressive image generation following key personnel departures from Meta.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Defining Agents and Classifying Autonomous Systems | 5 | 4 | 2 | 1 | Alessio asks Dharmesh for his definition of an agent. Dharmesh presents his deliberately broad framing that an agent is any AI-powered software accomplishing a goal, contrasting practical utility against overly theoretical agent frameworks. | |
| Atomic Agents, Tool Use, and Biological Metaphors | 5 | 5 | 2 | 1 | Swyx provides historical context on the agent winter and modern tool calling. Dharmesh introduces his atomic agent thesis using single-cell biological analogies, arguing tools are essentially primitive agents. | |
| Graph Theory, Distributed Consciousness, and Knowledge Graphs | 6 | 5 | 2 | 4 | Swyx raises the topic of distributed consciousness before the discussion shifts to knowledge graphs versus vector embeddings in RAG. Dharmesh elaborates on graph representations and his NodeRank concept, while Swyx notes that ML practitioners often view knowledge graphs with skepticism. | |
| Engineering Pragmatism, Abstraction Levels, and Vibe Coding | 6 | 5 | 2 | 5 | Swyx directly pushes back on the practicality of evaluating graph architectures. Dharmesh lays out his pragmatic philosophy favoring under-engineering over premature abstraction, and Alessio connects this to vibe coding and the risk of feature bloat. | |
| The Evolution of Engineers and The Power of MCP | 6 | 4 | 2 | 3 | Swyx asks whether junior engineers are obsolete, prompting Dharmesh to defend foundational systems thinking. The hosts and guest then evaluate Anthropic's Model Context Protocol (MCP) versus OpenAPI, with Dharmesh praising MCP's balance of simplicity and discoverability. | |
| Open Data Standards and the Open Graph Vision | 6 | 4 | 1 | 2 | Dharmesh outlines his vision for Open Graph data portability to break proprietary silos in platforms like LinkedIn. Swyx connects this concept to Bluesky's AT Protocol and Dan Abramov's decentralized identity models. | |
| Agent.ai Architecture: Building a Professional Network for Agents | 5 | 6 | 1 | 1 | Dharmesh breaks down the architecture and vision behind Agent.ai as a professional network and discovery registry for digital workers. Alessio shares his practical experience building a Latent Space researcher agent on the platform. | |
| Low-Code Workflows and the Future of Generative UI | 6 | 5 | 2 | 2 | Alessio inquires about deterministic low-code workflows versus LLM reasoning paths. Dharmesh explores generative UI caching, prompting Swyx and Alessio to compare sandbox environments like E2B and existing UI generation tooling. | |
| Agent Evaluation, Model Routing, and the Open Source Market | 6 | 5 | 2 | 2 | Alessio and Dharmesh discuss verifiable proof-of-work evals for hiring agents. Swyx asks about potential competition from OpenAI's GPT Store, leading Dharmesh to explain model routing and marketplace arbitrage. | |
| Business Models: Work as a Service vs. Results as a Service | 6 | 6 | 2 | 2 | Dharmesh explains why the software market is over-indexing on Results-as-a-Service, noting that objective metrics and predictable values exist in customer support but fail in subjective domains. Swyx validates this with an anecdote on 99designs logo contests. | |
| The Future Value of Software Engineers and Agent Memory Systems | 6 | 5 | 1 | 2 | Dharmesh presents a bullish case for software engineers by arguing total addressable economic value expands faster than automation. The conversation moves into cross-agent memory architectures and OAuth granular permissions, with Swyx cataloging current memory frameworks. | |
| Domain Investing Strategies and the Chat.com Deal | 4 | 6 | 1 | 1 | Swyx asks Dharmesh about his domain portfolio and domain acquisition tactics. Dharmesh shares the backstory behind acquiring chat.com and negotiating its transfer to Sam Altman, alongside his transparent, brokerless approach to domain purchasing. | |
| Founder Conviction, Information Diets, and Multimodal Innovations | 7 | 4 | 1 | 2 | Swyx inquires what drives Dharmesh's formidable conviction and how he consumes AI research. After Dharmesh highlights Gemini's recent image editing features, Swyx demonstrates deep technical domain knowledge by detailing the shift from diffusion models to autoregressive image generation. | |
| The 'Sorry Must Pass' Rule and the Future of AI Engineering | 5 | 5 | 0 | 1 | Alessio and Dharmesh discuss the philosophy of the 'Sorry Must Pass' rule for personal time management. Swyx connects Dharmesh's foundational work on inbound marketing to the nascent AI engineer movement, asking how to avoid common community failure modes. |
Statements from this episode (35)
Shah: Traditional software UIs are unintuitive compared to natural language
“As software developers, myself included, we've always said, oh, we build intuitive, easy to use applications. And it's not intuitive at all, right? Because what we're doing is taking the mental model that's in our head of what we're trying to accomplish with s…”
Shah: AI's next leap is autonomous multi-step action, not chat
“If we want to do something even more meaningful, it felt like the next kind of advancement is not this kind of, I'm chatting with some software in a kind of a synchronous back and forth model is that software is going to do things for me. In kind of multi-step…”
Shah: Early frameworks like AutoGPT assumed model reasoning that didn't exist
“So if you look at the first Implementation of like agent frameworks. You look at you know, baby AJI and auto GBT. I think it was an auto gen. That's the Microsoft one. They were way ahead of their time because they assumed this level of reasoning and execution…”
Shah: 2026 will be the year of multi-agent networks
“Lots of people have said it and you've hopefully combined some of those clips of really smart people saying this is the year of agents. And I completely agree. It is the year of agents, but then shortly after that is going to be the year of multi-agent systems…”
Shah: Graph representations may outperform standard vector chunking in RAG
“The reality is something gets lost in the chunking process in the, okay, well, those tend to, you know, like you don't really get the whole picture, so to speak, and maybe not even the right set of dimensions on the kind of broader picture. And it makes intuit…”
Shah: Graph Representations Offer Observability That Vector Embeddings Lack
“They're much more discoverable. You can kind of see it. There's observability to it versus kind of embeddings, which you can't really do much with as a human. You know, once they're in there, you can't pull stuff back out”
Swyx: Most ML practitioners consider knowledge graphs a 'dirty word'
“Most ML practitioners would say that knowledge graph is kind of like a dirty word the graph database. People get graph religion, everything's a graph, and then they go really hard into it, and then they get a graph that is too complex to navigate.”
Shah: Under-engineering is preferable because tech debt interest is quantifiable
“I would rather under-engineer something than over-engineer it if I were gonna err on the side of something. And here's the reason is that when you under-engineer it yes, you take on tech debt but the interest rate is relatively known and payoff is very, very p…”
Shah: AI codebase refactoring will make under-engineering even more optimal
“Because we're gonna, not that long from now, we're gonna have, you know, large code bases be able to exist you know, as context for a code generation or a code refactoring model. So, I think it's going to make it make the case for under engineering even strong…”
Shah: Zero-cost AI feature generation risks creating bloated, complicated software
“One of the risks that we have is that because adding a feature like a save or whatever the feature might be to a product, as that price tends towards zero, are we going to be less discriminated about what features we add as a result of making more product Prod…”
Shah: Systems thinking and abstractions retain timeless engineering value
“I think there's going to be timeless value in systems thinking and abstractions and what that means and whether functions manifested as math, which he's going to get exposed to regardless, or there are some core primitives to the universe. I think that the mor…”
Shah: MCP or an equivalent standard will be AI's next major unlock
“So I think MCP or something like it is going to be the next major unlock because it allows systems that don't know about each other, don't need to just decoupling of Like Sentry and whatever tools someone else was building.”
Shah: MCP provides distinct value above OpenAPI for LLM workflows
“I do think MCP as a pro adds value above open API. It's yeah, just because it solves this particular thing. And if we had come to the world, which we have, like, it's like, Hey, we already have open API. It's like, if that were good enough for the universe wou…”
Shah: Meta and LinkedIn keep social and professional graphs aggressively closed
“Right now our information, all of us nodes are in the social graph at Meta or the professional graph at LinkedIn, both of which are actually relatively closed and actually very annoying ways. Like very, very closed, right? Especially LinkedIn.”
Shah: Hybrid workplace teams of humans and AI agents are inevitable
“So I think it is, I will go so far as to say it's inevitable that we're going to have hybrid teams someday. And what I mean by hybrid teams. So back in the day, hybrid teams were, oh, well, you have some full-time employees and some contractors. Then it was li…”
Shah: Agent.ai has 1.3M users and 1,000 published agents
“So now, Agent.ai has 1.3 million users. 3000 people have actually, you know, built some variation of an agent, sometimes, sometimes just for their own personal productivity, about a thousand of which have been published.”
Shah: Deterministic steps beat LLM randomness when workflow steps are known
“Thing number two is if you can get, if you know in your head what the actual steps are to accomplish whatever goal, why would you leave that to chance? There's no upside. There's literally no upside. Just tell me like, what steps do you need executed?”
Shah: Agent interaction models will shift toward asynchronous queued workflows
“So we're used to the chat bot back and forth. Fine. I get that. I think we're gonna move to a blend of some of those things are gonna be synchronous as they are now, but some are gonna be async. It's just gonna put it in a queue”
Shah: The AI industry does not need a new language to replace Python
“He was talking about like, oh, we need a different language than Python or whatever that is like built for built for AI and built. It's like, No, Brett, I don't think we do actually. It's just fine. It deals with just fine, just expressive enough. And it's nic…”
Shah: Future AI will invent novel UI primitives beyond traditional controls
“Where AI is going to be headed on, I think, on the UI front is the same place as headed on the science front. That originally it's like, oh, well, based on the things that we know right now, it'll sort of combine them, but we're like right at the cusp of it be…”
Shah: Agent.ai automatically generates REST APIs and MCP integration for agents
“Every agent that's on agent.ai automatically has a REST API that's callable in exactly the way you would you'd expect. Automatically shows up in the MC, MCP server.”
Shah: OpenAI will inevitably build a custom AI agent marketplace
“I'm an investor, but no inside information. Is because it makes too much sense for them not to like, and they, they've taken multiple passes at it, right? They did the plugins back in the day, then the custom GPTs, and then the GPT store, because, you know, be…”
Shah: Model routing achieves multi-order-of-magnitude cost cuts without quality loss
“We can get a dramatic multiple words of magnitude reduction by going to a lower model with literally like no change in the quality of the output.”
Shah: The AI industry is over-indexing on outcome-based pricing
“I think the reason we're over indexed though, is that there are not that many use cases that have those two dimensions to them that are objectively measurable and that there's a known economic value that's constant, right?”
Shah: Web3 will make a comeback based on fundamental principles
“I think web three in the way that it was meant to be done. Is going to make a comeback because fundamental principles of that makes sense.”
Shah: Long-term economic value of software engineers will increase due to AI
“I think so I'm actually bullish on engineers in terms of their kind of long-term economic value.
Not despite all the movements in Cogen and all the things that we're, you know, already seeing, but because of it because what's going to happen as a result of…”
Shah: I built a personal vector store indexing 3 million of my emails
“So I have three million that I've built a vector store off of, did it solve my own personal use cases.”
Shah: Investor in OpenAI, Perplexity, LangGraph, CrewAI, and Limitless
“Investor in OpenAI, perplexity, lane graph, crew AI, limitless, a bunch of them.”
Shah: Agent.com will end up being more valuable than $15M Chat.com
“It's, yeah, it's gonna be, I think, end up being bigger than chat.com, which was 15.”
Shah: I bought crew.ai and offered it to CrewAI at cost
“By the way, I also own crew.ai, which I've offered, I'm an investor in. Yes. I bought that. And I've told him that like, whenever you're ready, you let me know, I'll sell it to you at cost.”
Shah: Owning premium domains grants access to competitive startup investment deals
“I owned playground.com... Suhail was out there with Playground the company... And he asked me whether I would consider... But once again, I took took equity. So it's like, I got the bright side. That's like, I, so domains that get me into deals that I would ne…”
Shah: I plan to stay at HubSpot for another 18 years
“So for those of you out there looking to kind of compete with HubSpot no I'm going to be here in 18 years, I'm going to be here for another 18 years.”
Shah: Conviction belongs on persistent problems, not specific products or solutions
“I don't generally tend to have conviction around a solution or a product. I have conviction around a problem that says this is an actual real problem. That needs to be solved. And I may have an idea for how to be solved you know, right now and that I may be ge…”