The Three-Layer AI Stack: Build, Operate, Remember

Greg Isenberg lays out a three-layer model for running a software business mostly on AI agents. The pitch is that you stop doing the grunt work yourself by stacking three things in order: a layer that builds the product, a layer that runs operations, and a layer that gives the agents memory. The video uses three boxes, color-coded red, green, and dark green, to walk through each layer and names a specific tool for each job.

Layer One: Build It With Claude Code

The first layer is explaining what you want to build with Claude Code. You describe the product and iterate with the AI until you have something finished. Isenberg claims the timeline should be under a week to go from idea to a working product. The visual for this layer is the red box, framed as the "build" piece of the stack. The point is that the build step is no longer the bottleneck, so it gets the least attention in the rest of the video.

Layer Two: Operate With Harbor

The second layer is operations, and the tool named is Harbor. The frames show Harbor as a node connecting multiple green dots with dashed lines, which represents connecting all your AI agents so they can delegate tasks to each other. You set up the agents, define the handoff rules, and then run it. The setup is shown as a checklist: "Set up the agents" and "Define the [handoff rules]."

The concrete example given is a support workflow that runs without you micromanaging it:

A support agent finds a critical bug.
It tells the dev agent to ship the fix.
A third agent closes the ticket and emails the user.

The "agents" frame shows a hub-and-spoke constellation, one central node wired out to many smaller nodes, which is the mental model for one orchestration layer coordinating many specialized agents.

Layer Three: Memory, The Part Most People Skip

Isenberg's main argument is that most people stop after the ops layer, and that is the mistake. His line: agents can only act on what they know. Without a memory layer, the agents are running blind on whatever is in front of them. The dark green box is the "memory layer."

He breaks the memory layer into three steps, shown as a numbered list ending in the Pinecone logo:

Record every customer call using Fireflies.ai.
Pull the transcripts using a cron job. (The transcript says "crown job," which is a speech-to-text error for cron job, a scheduled task.)
Index everything with Pinecone, a vector database.

So calls get recorded, transcripts get pulled on a schedule, and everything gets indexed into a searchable vector store the agents can query. That indexed memory is what lets the ops agents act on real context instead of starting from zero each time.

The Full Stack

Put together, the flow is: Claude Code builds the product, Harbor wires the agents together to run operations, and the Fireflies-to-cron-to-Pinecone pipeline feeds the agents memory. The closing visual is two hands joining puzzle pieces with the caption "that's the last piece of the puzzle," framing memory as the missing component that turns a collection of agents into a system that runs itself.

Key Takeaways

The stack has three layers in order: build (Claude Code), operate (Harbor), and remember (Fireflies + cron + Pinecone).
Claude Code handles the build layer, with a claimed timeline of under a week from idea to finished product.
Harbor is the orchestration layer that connects agents and lets them delegate tasks to each other based on handoff rules you define.
A worked example: a support agent finds a bug, hands it to a dev agent to ship the fix, and a third agent closes the ticket and emails the user, all without manual oversight.
The memory layer is the step most people skip, and Isenberg argues it matters most because agents can only act on what they know.
Memory is built in three steps: record calls with Fireflies.ai, pull transcripts on a cron schedule, and index everything into Pinecone as a vector database.