HubSpot co-founder Dharmesh Shah charts the pragmatic evolution of AI agents, revealing his journey from 20-year-old natural language concepts to the current state of agent development and the imminent rise of multi-agent networks, offering critical insights for Crypto AI investors and researchers navigating this rapidly advancing field.
Dharmesh Shah's Journey into AI and Natural Language
- Dharmesh Shah, introduced by hosts Alesio and Swyx, recounts his long-standing fascination with natural language interfaces for software, dating back 20 years to a pre-HubSpot concept called Innosoft. Innosoft aimed to use email as an offline-capable, natural language interface for business software like CRMs, parsing commands to update systems. While ahead of its time, this core idea resurfaced post-ChatGPT with projects like ChatSpot, driven by the realization that modern AI finally makes effective natural language interaction with software feasible. Dharmesh emphasizes the shift from clunky UI interactions to simply expressing intent in natural language as a fundamental breakthrough.
- Strategic Insight: The long gestation period of natural language interfaces highlights that foundational concepts in AI often predate viable technology; investors should look for historical ideas now becoming feasible due to advancements like LLMs.
Defining AI Agents: A Pragmatic and Broad View
- Asked for a definition, Dharmesh offers a deliberately broad and potentially "irritating" take: an agent is simply "AI-powered software that accomplishes a goal." He acknowledges the criticism of this breadth but argues it encompasses the diverse implementations emerging.
- Dharmesh suggests classification (autonomous vs. non-autonomous, deterministic vs. non-deterministic workflow, sync vs. async) is more useful than a narrow definition, viewing the agent landscape as overlapping categories within this broad umbrella. His focus remains pragmatic, centered on what can be built now to solve real problems with reliability.
- Quote: "I have a very broad definition for agent... it's AI powered software that accomplishes a goal period. That's it." - Dharmesh Shah
- Actionable Implication: The lack of a single, narrow definition suggests the agent space is still formative. Researchers should focus on classifying agent capabilities and limitations, while investors should look for practical applications solving specific goals rather than betting on one specific agent architecture winning out immediately.
Critiquing Early Agents and the "State of the Practical"
- Dharmesh reflects on early agent frameworks like BabyAGI and AutoGPT, viewing them as valuable thought experiments that were ahead of their time, assuming reasoning and planning capabilities that didn't yet exist reliably.
- He contrasts the "state-of-the-art" with his preferred focus: the "state of the practical"—what can actually be built and deployed effectively today to solve discrete problems with verifiable results. This pragmatic lens, shaped by his extensive software engineering background, prioritizes current feasibility over purely theoretical potential.
- Crypto AI Relevance: This mirrors challenges in crypto adoption; focus on practical, verifiable use cases (like specific DeFi protocols or verifiable compute) often yields more near-term value than overly ambitious, complex systems that lack robust infrastructure or clear market fit.
The "Why Now" for Agents: Enabling Factors
- Swyx outlines key factors driving the current "summer" of agents after an initial "winter": better models, more reliable tool use (like mCP - Multi-Context Prompting, a protocol allowing AI models to interact with external tools and data sources more effectively), shifting business models (like Results-as-a-Service), dropping costs, faster inference, and greater model diversity.
- Dharmesh agrees these factors are converging, making previously theoretical agent concepts increasingly practical.
- Investor Takeaway: The confluence of these factors (improving models, better tooling like mCP, cost reduction) creates fertile ground. Monitor progress in each area, as bottlenecks shifting (e.g., from model capability to tool reliability or cost) signal new investment or research opportunities.
Atomic Agents, Tool Use, and Multi-Agent Systems
- Dharmesh introduces his concept of "Atomic Agents"—the simplest possible AI-powered unit analogous to a single-celled organism.
- He provocatively suggests that current "tool use" in LLMs could be reframed as one agent delegating tasks to specialized, atomic agents (even if those "tools" are simple functions). This "turtles all the way down" perspective simplifies the architecture to a network of collaborating agents, potentially paving the way for more complex multi-agent systems, which he predicts will be the major focus after the "year of agents."
- Technical Definition: Tool Use/Calling: The capability of Large Language Models (LLMs) to interact with external software functions or APIs to retrieve information or perform actions beyond their internal knowledge.
- Strategic Consideration: Viewing tools as agents suggests a future of highly composable AI systems. For Crypto AI, this could mean decentralized networks where specialized, verifiable "atomic agents" (perhaps running on-chain or via zkML) perform specific tasks within a larger workflow.
Graphs as Knowledge Stores for AI
- The conversation shifts to data structures, specifically graph databases as potential knowledge stores for AI, contrasting them with relational, document, or vector databases.
- Dharmesh, citing his long interest in graph theory (like PageRank), posits that graphs might represent knowledge more effectively for LLMs than chunked vectors from RAG (Retrieval-Augmented Generation) – a technique enhancing LLM responses by retrieving relevant information from external knowledge bases.
- While acknowledging the practical challenges and potential complexity ("graph religion"), he sees promise in their ability to capture relationships and context potentially lost in other methods.
- Crypto AI Angle: Knowledge graphs could be crucial for mapping complex relationships in decentralized systems (e.g., token flows, governance interactions, protocol dependencies). Research into efficient, verifiable graph representations for AI could unlock new analytical capabilities for on-chain data.
Engineering Philosophy: Pragmatism and Managing Complexity
- Dharmesh shares his engineering philosophy, emphasizing pragmatism and return on investment (or "return on calories").
- He generally prefers under-engineering to over-engineering, arguing that the cost of fixing under-engineered solutions (technical debt) is often more predictable and manageable than the cost of building unnecessary complexity upfront, especially as AI-driven refactoring tools improve.
- He advocates for doing things "the right way" only when the marginal cost is low, focusing on solving the immediate problem effectively.
- Insight for Builders: In the fast-moving Crypto AI space, this pragmatic approach is vital. Prioritize shipping functional components and iterating, rather than attempting perfect, overly complex systems upfront. The cost of future refactoring may decrease with better AI tooling.
mCP: A Key Standard for Interoperability
- Dharmesh expresses strong enthusiasm for mCP (Multi-Context Prompting), viewing it as a crucial, practical standard enabling interoperability between AI systems and tools.
- He sees it as superior to generic standards like OpenAPI for the specific use case of LLM tool discovery and interaction because it's "simple enough to both be useful and understandable."
- He believes mCP (or similar standards) is a major unlock for multi-agent systems, allowing agents to discover and delegate tasks to each other without tight coupling, much like microservices interact via APIs.
- Crypto AI Opportunity: Standards like mCP are vital for decentralized AI. Investors and researchers should track mCP adoption and development, as it could form the basis for interoperable protocols enabling communication and collaboration between agents across different platforms or even blockchains.
Data Ownership, Standards, and "Open Graph"
- The discussion touches on data silos (using LinkedIn's closed nature as an example) and Dharmesh's unrealized idea for "Open Graph"—a standard allowing individuals to own and openly publish their professional or social graph data in a controllable, opt-in manner.
- He contrasts this with the current reality where user data is locked within platforms. This ties into the need for better data access and control, potentially enabled by new standards or even Web3 principles focused on user ownership and verifiable data.
- Web3 Connection: Dharmesh explicitly notes that the principles of Web3 (verifiable data, audit logs, user control) are relevant for solving attribution and data ownership challenges, even if past implementations were flawed. This reinforces the potential synergy between AI agent needs and decentralized technologies.
Agent.ai: A Professional Network for AI Agents
- Dharmesh elaborates on his current project, Agent.ai, conceived as a "LinkedIn for Agents." The core thesis is that AI agents will increasingly function as "teammates" alongside humans in hybrid teams.
- To facilitate this, agents need a professional network for discovery, capability assessment, and interaction. Agent.ai aims to provide profiles, connections, and potentially even a low-code platform for building agents, fostering a collaborative ecosystem where agents can leverage each other's capabilities via standards like mCP.
- Market Trend: The emergence of platforms like Agent.ai signals a move towards structured ecosystems for agent discovery and collaboration. Crypto AI projects might consider similar registry/discovery mechanisms for decentralized agents or AI-powered services.
Agent Composition, UI, and Memory
- The conversation explores the practicalities of building agents on platforms like Agent.ai, including the challenge of composing agents (how many tools/agents can an LLM handle?) and the future of user interfaces (moving beyond simple chat to asynchronous tasks and potentially AI-generated UIs).
- Dharmesh highlights the critical need for long-term memory in agents, not just for individual agents but also cross-agent memory where knowledge learned by one agent about a user can be securely shared (with consent) with other agents to avoid redundant interactions.
- He also emphasizes the need for fine-grained authorization controls beyond current oAuth scopes.
- Research Frontier: Cross-agent memory and fine-grained, user-controlled authorization are critical unsolved problems. Solutions leveraging cryptographic techniques (like ZK proofs for privacy-preserving data sharing) or decentralized identity could be highly valuable in the Crypto AI space.
Business Models: Work-as-a-Service vs. Results-as-a-Service
- Dharmesh distinguishes between Work-as-a-Service (software performing tasks, priced per use/time) and Results-as-a-Service (priced per outcome).
- While acknowledging the appeal of outcomes-based pricing (popular in customer support AI), he cautions that it's only suitable for use cases with objectively measurable outcomes and relatively stable economic value per outcome.
- Many tasks, especially creative or complex ones, don't fit this model well, suggesting Work-as-a-Service (or traditional SaaS) will remain dominant for many agent applications.
- Investor Note: Evaluate AI business models critically. Results-as-a-Service sounds appealing but faces significant measurement and valuation challenges. Work-as-a-Service or hybrid models may be more broadly applicable, especially in enterprise or complex B2B scenarios relevant to many crypto/blockchain applications.
Conviction, Learning, and Saying No
- Reflecting on his drive and decision-making, Dharmesh attributes his "fierceness" to conviction (often focused on problems, not specific solutions), a willingness to grind, and a long-term perspective.
- He emphasizes the importance of continuous learning (mentioning YouTube as a key resource) and the necessity of saying "no" to maintain focus, referencing his "Sorry Must Pass" philosophy to manage overwhelming inbound requests without guilt.
- Personal Strategy: For researchers and investors in the noisy Crypto AI field, developing conviction around core problems (e.g., decentralized compute, verifiable AI, data privacy) and ruthlessly prioritizing focus are essential survival skills.
Dharmesh Shah underscores the shift towards practical, interconnected AI agents, moving beyond theoretical concepts. Crypto AI investors and researchers must track evolving agent frameworks, interoperability standards like mCP, and memory solutions, as these developments signal critical infrastructure needs and opportunities within decentralized AI ecosystems.
