The Agent Network — Dharmesh Shah

Episode Summary

Executive Summary: Dharmesh Shah argues that agents are simply AI-powered software that accomplishes goals, and that the real shift is toward multi-agent networks, shared memory, and standards like MCP that let agents discover, delegate, and compose work. He connects this to practical product design, code generation, model routing, and new business models where software increasingly does the work itself.

Main Topics: Defining agents pragmatically (Priority: 5/5): Dharmesh offers a broad, goal-based definition of agents as AI-powered software that accomplishes a goal, while acknowledging the term is overloaded and best understood through useful classifications like autonomy, workflow type, and interaction mode. From chat UX to multi-agent networks (Priority: 5/5): He frames chat as an important but limited interface and argues the next step is agents that do multi-step work asynchronously, eventually forming multi-agent systems and digital teams. MCP, discovery, and composability (Priority: 5/5): Dharmesh sees MCP as a major unlock because it standardizes discovery and tool access for LLMs and agents, enabling a network of interoperable capabilities rather than isolated integrations. Agent.ai as a marketplace and professional network (Priority: 5/5): He describes Agent.ai as a free agent builder evolving into a marketplace and professional network for AI agents, with profiles, reviews, shared memory, and composable agents discoverable through APIs and MCP. Memory, auth, and shared state (Priority: 4/5): A major frontier is long-term memory across agents and teams, with selective authorization and fine-grained sharing so users can reuse context without exposing everything. Code generation, UI generation, and vibe coding (Priority: 4/5): He argues that lower code-generation costs will push more under-engineering, but also require better guardrails against product bloat; he expects UI generation to become a stable, cached artifact in agent workflows. Business models, attribution, and market efficiency (Priority: 4/5): Dharmesh distinguishes work-as-a-service from results-as-a-service, arguing outcomes-based pricing works best where value is measurable and markets are efficient, but many domains still lack attribution and reliable evaluation.

Key Arguments: Agents should be defined broadly as AI-powered software that accomplishes a goal, because the ecosystem contains many overlapping agent types rather than one canonical form. The practical frontier is not just better models but better orchestration: tool use, MCP, model routing, memory, and discovery are what make agents commercially useful. Multi-agent systems will follow the current wave of single agents, because real work is naturally decomposable and collaborative. MCP matters because OpenAPI alone does not solve LLM-specific discovery and delegation; a lightweight standard can unlock interoperability across tools and agents. Agent.ai is intended to be a network effect product: agents, users, memory, reviews, and MCP exposure all reinforce one another. Shared memory across agents and teams will create more user value than isolated per-agent memory, but it requires selective, trustable authorization. As code generation gets cheaper, the risk shifts from under-engineering to over-adding features and complexity; product discipline becomes more important, not less. Results-based pricing only works well when outcomes are objectively measurable and economically comparable, which is true in some support-like workflows but not many subjective domains. Engineers remain valuable because AI expands the total problem space faster than it commoditizes engineering labor. The future of software may include AI-generated UI and even new interaction primitives beyond today’s text boxes, checkboxes, and dropdowns.

Data Points: chat.com sale price: $15 million - Dharmesh references the sale of the chat.com domain to OpenAI as a major prior milestone. Agent.ai users: 1.3 million - He says Agent.ai has reached this user count. Agents built on Agent.ai: 3,000 - Number of people who have built some variation of an agent on the platform. Published agents on Agent.ai: about 1,000 - Subset of built agents that have been published publicly. Agent.ai reviews: tens of thousands - He cites review volume as a signal for agent quality and evaluation. Average agent rating: 4.1/5 stars - Average rating across existing agents on Agent.ai. Domain valuation example: $25,000 vs $5,000 - He describes an arbitrage use case where an agent estimates a domain at $25k while it is listed for $5k. Model selection behavior: GPT 4.5 chosen most often - He says users tend to pick the highest-numbered model rather than auto-optimization. Email backlog: 3 million emails - He mentions using email as a life data bus and building a vector store over millions of emails. Typical bedtime: 2 a.m. - He describes his late-night work and YouTube/coding routine.

Pivotal Quotes: "AI-powered software that accomplishes a goal. Period." — Dharmesh Shah: His core definition of an agent. "I think it's inevitable that we're going to have hybrid teams someday." — Dharmesh Shah: He predicts teams composed of humans and AI agents working together. "The thing I worry about is that if it starts to become too easy, are we going to be too promiscuous in our kind of extension, adding product extensions and things like that?" — Dharmesh Shah: He warns that cheaper code generation may lead to bloated products.

Implications: The conversation suggests agents are moving from demos to infrastructure: standards, memory, routing, and marketplaces will matter as much as model quality. Builders should focus on composability, evaluation, and selective trust, while users should expect AI to become a teammate rather than just a chatbot.