Key Moments
The Agent Network — Dharmesh Shah, Agent.ai + CTO of HubSpot
Latent Space PodcastKey Moments
Dharmesh Shah discusses AI agents, defining them broadly, and their future potential, alongside graph theory, engineering practices, and business models.
Key Insights
AI agents are broadly defined as AI-powered software that accomplishes a goal, with future development likely leaning towards multi-agent systems.
Graph theory and knowledge graphs offer a more elegant and discoverable way to represent knowledge for AI compared to traditional databases.
Over-engineering can be detrimental; a pragmatic approach is to do things the 'right way' when the cost is marginal, and otherwise, err on the side of under-engineering to avoid unnecessary debt.
MCP (Message Communication Protocol) is a crucial standard for agent communication and collaboration, unlocking new possibilities for multi-agent systems.
The future likely involves 'hybrid digital teams' composed of humans and AI agents, necessitating professional networks for agents similar to LinkedIn for humans.
While LLMs are improving, challenges remain in areas like memory, UI generation, and precise arithmetic, but access to tools like Python interpreters is significantly enhancing capabilities.
DEFINING AND EVOLVING AI AGENTS
Dharmesh Shah offers a broad definition of AI agents as "AI-powered software that accomplishes a goal." He acknowledges the criticism of this definition's breadth but argues it's useful for understanding future developments, especially in multi-agent systems. He traces the evolution from early attempts at natural language interfaces to current capabilities, highlighting the importance of AI agents for accomplishing complex, multi-step tasks.
THE ROLE OF GRAPH THEORY AND KNOWLEDGE REPRESENTATION
Shah expresses fascination with graphs and graph theory, viewing them as powerful tools for knowledge representation. He suggests that knowledge graphs could be a superior alternative to traditional databases or even vector stores for LLMs, offering better structure and discoverability. Concepts like PageRank are being adapted to rank nodes within these graphs, adding another layer of sophistication for AI applications.
PRAGMATIC ENGINEERING AND THE "SORRY, MUST PASS" PHILOSOPHY
Discussing engineering practices, Shah advocates for pragmatism, emphasizing doing things the right way when the incremental cost is minimal. He stresses the importance of return on investment for engineering efforts and favors under-engineering over over-engineering to avoid potential future debt. His "Sorry, Must Pass" philosophy, born from being overwhelmed, highlights the need to say no to opportunities that don't align with core convictions, emphasizing focus and preventing burnout.
THE RISE OF MCP AND COLLABORATIVE AGENTS
MCP (Message Communication Protocol) is hailed as a significant advancement, providing a much-needed standard for agent communication and discovery. Shah sees MCP as foundational for multi-agent systems, enabling agents to collaborate and delegate tasks effectively. He envisions a future where agents can discover each other's capabilities through registries, leading to powerful compositions of specialized agents.
THE EMERGENCE OF HYBRID DIGITAL TEAMS AND AGENT NETWORKS
Shah foresees the inevitable rise of 'hybrid digital teams,' comprising both humans and AI agents. This evolution necessitates the creation of professional networks for agents, akin to LinkedIn for humans, where agents have profiles, post updates, and collaborate. Agent.ai is presented as an example of such a platform, aiming to build a foundational network for these future automated workers.
BUSINESS MODELS AND THE FUTURE OF WORK
The conversation explores 'Work as a Service' where software performs tasks, and 'Results as a Service,' charging for outcomes. While customer support is a prime example of the latter due to measurable metrics, Shah suggests other use cases might be better suited for 'Work as a Service' due to subjectivity. He also discusses how AI will augment, not replace, engineers, increasing their value by enabling them to solve more complex problems.
MEMORY, UI GENERATION, AND MODEL SELECTION
Long-term memory, especially cross-agent memory, is identified as a critical frontier, enabling agents to build on past learnings and user interactions. Challenges in UI generation are being addressed by code generation models, potentially simplifying interface creation. Furthermore, efficient model selection and routing are becoming crucial as users often over-rely on the most powerful, costly models, highlighting an opportunity for optimization and cost reduction.
THE IMPORTANCE OF HAVING CONVICTION AND SOLVING PROBLEMS
Shah's formidable nature stems from a strong conviction in problems rather than solutions. He believes in latching onto a real problem and revisiting it as technology evolves, rather than pursuing a specific product. This patient, problem-centric approach allows him to adapt and persist through technological shifts, ultimately driving innovation and building enduring companies.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Concepts
●People Referenced
Common Questions
Dharmesh Shah defines an AI agent as "AI-powered software that accomplishes a goal." He acknowledges this is a very broad definition but argues it encompasses the diverse implementations and types of agents emerging, from autonomous to non-autonomous, and deterministic to non-deterministic workflows.
Topics
Mentioned in this video
A proposed standard for an open, user-controlled professional graph, allowing individuals to publish and control access to their data, distinct from Meta's Open Graph.
An authorization framework used for granting access to data, but criticized for its slow evolution and lack of fine-grained control over data sharing.
A type of database used for storing vector embeddings, mentioned as an alternative to traditional and graph databases for knowledge representation in AI.
A standard that enables disparate systems to communicate and collaborate, seen as a crucial unlock for multi-agent systems, allowing discovery and delegation.
A standard for describing RESTful APIs, which MCP is argued to marginally improve upon for LLM tool discovery.
A company Dharmesh Shah contemplated 20 years ago, focused on natural language interfaces to business software, which was ahead of its time.
Dharmesh Shah's current passion project, a professional network for AI agents, aiming to foster collaboration and discovery among agents.
A domain registrar where domain names are listed for sale, mentioned in the context of domain valuation.
A graphic design platform also facing potential disruption from advanced AI image generation and editing tools.
A project mentioned for its contributions to agent memory, highlighting the ongoing development in this area.
A language model considered as an alternative to Mistral for specific steps in agentic workflows.
The AI model that made natural language interfaces viable for Dharmesh Shah's long-held vision, enabling structured text conversion with few-shot examples.
An early agent framework, along with BabyAGI, that was ahead of its time, assuming a level of reasoning and execution planning capability that didn't yet exist.
An early agent framework that, while not fully ready, signaled the future of agent capabilities.
An AI framework whose domain name, Crew.ai, was acquired by Dharmesh Shah and intends to offer to the company at cost.
A web analytics service from Google that can be connected to Google Search Console to access keyword data.
A domain name acquired by Dharmesh Shah with the idea of building a consumer-friendly chat product, later sold to OpenAI when ChatGPT became a full product.
A library for building language agent systems as graphs, mentioned for its hierarchy of memories.
A code generation tool used by Dharmesh Shah, enabling 'vibe coding' and auto-save functionalities.
An application used for app generation, mentioned alongside Bolt and Replit Agent.
A platform for freelance graphic design that creates an efficient market for logo design by connecting many designers with customers, described as fantastic by the speaker.
A Microsoft agent framework, distinct from earlier projects like BabyAGI and Auto-GPT.
A professional social media platform whose data is described as "relatively closed", highlighting the need for open data standards like Open Graph.
Dharmesh Shah's blog post and website explaining his philosophy of saying 'no' to most requests to avoid guilt and manage his time effectively.
The programming language of choice for Dharmesh Shah and deemed the 'lingua franca' of AI languages due to its expressiveness and utility for both humans and AI.
A company invested in by Dharmesh Shah, which provides a code sandbox and powered the LLM Arena web Arena for UI generation.
A platform for UI generation powered by e2b's code sandbox.
An application used for app generation, mentioned alongside Loveable and Bolt.
An open-source tool that uses a graph database for memory, highlighted as a notable project in the agent memory space.
A language model considered as an alternative to Claude for specific steps in agentic workflows.
An earlier version of ChatGPT capable of decent few-shot example conversions to structured text, instrumental in proving the viability of natural language interfaces.
Founder of Agent.ai and co-founder and CTO of HubSpot, known for his long-standing interest in natural language interfaces and AI agents.
Formerly a prominent figure in React development, now works on Blue Sky and advocates for user data ownership.
An entrepreneur known for his 'hell yes or no' philosophy on commitments, which aligns with Dharmesh Shah's approach to managing his schedule.
CEO of OpenAI, described as a 'fierce nerd' and 'formidable person,' with whom Dharmesh Shah avoids direct competition.
A well-known marketing and sales software company co-founded by Dharmesh Shah, which started as a project rooted in natural language interfaces.
An open-source framework for developing applications powered by language models, mentioned in the context of memory solutions for agents.
An application used for app generation, mentioned alongside Loveable and Replit Agent, and a contributor to generative UI.
A traditional image editing software that emerging autoregressive AI image generation models pose a potential threat to.
More from Latent Space
View all 92 summaries
86 minNVIDIA's AI Engineers: Brev, Dynamo and Agent Inference at Planetary Scale and "Speed of Light"
72 minCursor's Third Era: Cloud Agents — ft. Sam Whitmore, Jonas Nelle, Cursor
77 minWhy Every Agent Needs a Box — Aaron Levie, Box
42 min⚡️ Polsia: Solo Founder Tiny Team from 0 to 1m ARR in 1 month & the future of Self-Running Companies
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Try Summify free