Dharmesh Shah — Agent Network

In one sentence

The future of work lies in "hybrid teams" where human workers and AI agents collaborate as peers, necessitating a new infrastructure for agent discovery, professional networking, and shared memory.

Why read this (and who it's for)

This conversation explores the transition from "Chat" (synchronous interaction with a tool) to "Agents" (asynchronous, goal-oriented teammates). It is essential for AI engineers building multi-agent systems, product managers designing the next generation of work software, and founders trying to navigate the shift from SaaS to "Work as a Service" (WaaS). You will learn how to think about agentic architecture, the importance of "shared memory," and why the professional network for agents is the next major frontier.

The big idea

Dharmesh Shah argues that we are moving beyond the "Chat" era—where a user interacts with a model in a synchronous back-and-forth—into the era of Agents. An agent is defined simply as "AI-powered software that accomplishes a goal." Unlike a chatbot, an agent is designed to handle multi-step, often asynchronous tasks.

The core thesis is that these agents will eventually become teammates. Shah envisions a "hybrid team" model similar to how we transitioned from full-time to contract workers, or in-office to remote workers. In this model, agents aren't just utilities; they are entities with specific roles that require a "professional network" (Agent.ai) to discover, a standard for communication (MCP), and a shared memory to collaborate effectively.

Key ideas

The Definition of an Agent

• What it is — An agent is "AI-powered software that accomplishes a goal. Period." It is defined by its ability to handle non-deterministic workflows and act asynchronously.
• Why it matters — By stripping away the fluff, we can categorize agents by their interaction modes: Chat (synchronous), Workflow (discrete steps), and Autonomous (independent goal-seeking).
• In the book — Shah notes that while his definition is "broad as to be completely non-useful" to some, it is the only way to capture the full spectrum of what is becoming possible as models become more reliable.

Hybrid Digital Teams

• What it is — The evolution of the workplace where "carbon-based life forms" and AI agents work together as a unified team.
• Why it matters — This shifts the mental model of AI from a "tool" (like a hammer) to a "teammate" (like a junior developer or a researcher).
• In the book — Shah compares this to the evolution of hybrid work (remote vs. office). If agents are teammates, we need a way to "hire" them, which leads to the need for a professional network for agents.

Work as a Service (WaaS) vs. Results as a Service (RaaS)

• What it is — WaaS is charging for the work performed (e.g., "review this contract"), while RaaS is charging for the outcome (e.g., "X dollars per resolved ticket").
• Why it matters — While RaaS is popular in customer support because outcomes are easily measured, Shah argues we are "over-indexing" on it. Many AI applications don't have a clean, measurable "result" that carries consistent economic value per transaction.
• In the book — Shah argues that for many complex tasks, WaaS is a more honest and sustainable way to monetize AI capabilities.

The Professional Network for Agents (Agent.ai)

• What it is — A platform (LinkedIn for agents) where agents can be discovered, reviewed, and audited.
• Why it matters — As agents become teammates, we need a way to see what an agent is capable of, what its "experience" is, and what its "ratings" are.
• In the book — Shah envisions a world where a human can "hire" five different agents to solve a problem, run an "eval" on them, and pick the best one based on verifiable performance.

Shared Memory and Context

• What it is — The ability for agents to share state, context, and "memories" across different interactions and teams.
• Why it matters — For a hybrid team to work, agents need a "shared team" memory (like a Slack channel) rather than just isolated, individual memories.
• In the book — Shah discusses the importance of "cross-agent memory" and the need for granular, selective data sharing (similar to OAuth but with finer control).

The Role of MCP (Model Context Protocol)

• What it is — A standard for agent collaboration, tool use, and discovery.
• Why it matters — It allows systems that don't know about each other to communicate. It is the "decoupling" layer that allows for a registry of tools and servers.
• In the book — Shah argues that while OpenAPI is great for describing APIs, MCP is "use-case specific" and solves the unique problem of LLM discovery and tool-use.

Frameworks & tools

The Hybrid Team Framework

To transition from "Tool-based AI" to "Team-based AI," evaluate your workflow using these three layers:

1. Role Definition: What specific "job" is the agent doing? (e.g., Researcher, Coder, Reviewer).
2. Interaction Mode: Is the task Synchronous (Chat) or Asynchronous (Queue/Task)?
3. Collaboration Layer: Does the agent need to share memory with other agents or humans? (Shared Team State).

The WaaS vs. RaaS Decision Matrix

When deciding how to monetize or measure an agent's value, ask:

1. Is the outcome easily measurable? (e.g., "Ticket Resolved"). If yes, consider Results as a Service (RaaS).
2. Is the value in the process? (e.g., "Legal Review"). If the outcome is subjective or complex, use Work as a Service (WaaS).
3. Can we establish a baseline? If you can't define what a "success" looks like in a way that customers agree on, stick to WaaS.

Memorable examples & stories

• The "Offline" Email Idea: Shah recalls an idea from 20 years ago to use email as a natural language interface to business software because it worked offline. He notes that we are finally seeing the "natural language" breakthrough he envisioned.
• The "Junior Engineer" Crisis: Shah discusses how the value of an engineer is changing. While the "denominator" (number of engineers) might increase due to AI, the "numerator" (total economic value possible) is growing even faster because of the power tools AI provides.
• The "Vibe Coding" Calculus: Shah describes "vibe coding" as a way to lower the cost of "over-engineering." Because the cost of refactoring code is trending toward zero, it is often better to "vibe" a solution now and pay the "interest" (refactoring) later.

Key quotes

• "I have a very broad definition for agents, which is it's AI-powered software that accomplishes a goal. Period."
• "The next form of hybrid is like the carbon-based life forms and agents and AI and some form of software."
• "I think the value of an engineer actually goes up... they're going to have the power tools, they're going to be able to solve a larger base of problems."
• "Every time you say yes to anything, you are by definition saying no to something else." (Paraphrased) — This refers to the importance of saying "no" to maintain focus.

How to apply it

This week

• Audit your "Chat" interactions: Identify one task you currently do with a chatbot that could be a "multi-step" goal. Can you turn it into a "Workflow" where the agent does the work and notifies you when done?
• Evaluate your AI pricing: If you are building an AI tool, map out your current model. Is it RaaS or WaaS? If it's RaaS, ask: "Is this outcome actually easy for the customer to measure?"
• Practice "Vibe Coding": On a non-critical internal task, intentionally "under-engineer" the solution. Use AI to generate a "good enough" UI or function quickly, and only refactor if it actually breaks.

Ongoing

• Think in "Hybrid Teams": When planning a project, ask: "Which of these tasks can be delegated to a digital teammate (agent)?"
• Prioritize "Shared Memory": When building agentic systems, think about how the agent will "remember" the context of the team, not just the individual user.

Common pitfalls

• Over-indexing on RaaS: Many founders think they need to charge "per ticket" or "per result," but if the result isn't clearly defined, this leads to a poor user experience and hard-to-measure success.
• Over-engineering for the future: Don't spend 10 hours building a perfect, scalable architecture for a feature that might be replaced by a "vibe-coded" solution in a week.
• Anthropomorphizing: While we call them "teammates," they are still software. They should self-disclose as software to maintain trust and clarity.

Connects to

• "Sorry, Must Pass" by Dharmesh Shah: For the philosophy of saying "no" to protect your focus.
• The concept of "Multi-Agent Systems" (MAS): This conversation expands on the technical and organizational implications of MAS, specifically regarding professional networks and shared memory.

If you remember only three things

• Hybrid Teams: The future of work is a mix of humans and AI agents acting as teammates.
• WaaS vs. RaaS: Don't assume "Results as a Service" is the best model; "Work as a Service" is often more appropriate for complex, non-discrete tasks.
• Agentic Infrastructure: The next big unlocks are MCP (for discovery/standards) and Professional Networks (for trust and verification).

Internalize it

• Reflection questions
  • If I were to hire an AI agent today, what specific "job description" would I give it?
  • Which of my current daily tasks are "Chat" (synchronous) and which could be "Work" (asynchronous)?
  • How would my team's dynamic change if I added three "digital teammates" to our Slack channel?
• Flashcards
  • Q: How does Dharmesh Shah define an "Agent"?
  • A: AI-powered software that accomplishes a goal.
  • Q: What is the difference between WaaS and RaaS?
  • A: WaaS (Work as a Service) charges for the work performed; RaaS (Results as a Service) charges for the outcome.
  • Q: What is the "Hybrid Team" concept?
  • A: A workplace organization where human workers and AI agents collaborate as team members.
  • Q: Why is MCP important for agents?
  • A: It provides a standard for discovery and tool-use, allowing systems that don't know each other to collaborate.
  • Q: What is the "Professional Network" for agents?
  • A: A platform (like Agent.ai) where agents can be discovered, reviewed, and verified by their capabilities and "ratings."

Fidelity note

This summary is grounded in the provided transcript. I am confident in the definitions of Agents, Hybrid Teams, and the WaaS/RaaS distinction. The specific details regarding "Agent.ai" and "MCP" are based on the guest's personal projects and current industry trends discussed in the interview. The "Vibe Coding" section is a synthesis of his comments on "under-engineering" and "refactoring" costs.