Greg Isenberg hosts Remy Gaskill for a beginner-friendly full course on building AI agents. Remy draws a clear line between chat models (question-to-answer) and agents (goal-to-result), then walks through live demos showing Claude Code, Codex, and Antigravity completing tasks autonomously through an observe-think-act loop. The session covers the full agent stack: agents.md context files for persistent onboarding, memory.md for preference learning across sessions, MCP integrations for connecting real tools like Gmail and Notion, and skill files as reusable SOPs. By the end, Remy has demonstrated a working executive assistant that drafts emails, creates Stripe payment links, and sets up Notion projects from a single natural-language command -- and explains how compounding skills can eventually automate entire business departments.
Key Points
The core distinction is: chat models are question-to-answer (ping-pong), while agents are goal-to-result -- they plan, execute, and deliver autonomously without you babysitting each step.
Every agent operates on a three-step loop: observe (take in context and current state), think (decide the next action), and act (execute) -- cycling until the task parameters are satisfied.
An agent is made up of four components: an LLM as the brain, the loop for continuous execution, tools for interacting with the world, and context that shapes its behavior.
All agent harnesses -- Claude Code, Codex, Antigravity, Co-work, Open Claw -- are just applications that facilitate this loop. Once you understand the loop, you can operate any of them.
Unlike chat models, agents do not have automatic cloud memory. Context must be explicitly loaded via an agents.md (or Claude.md) file that acts as a persistent system prompt at every session start.
A memory.md file paired with an instruction in agents.md allows the agent to save and recall preferences across sessions, creating a self-improving loop that compounds over weeks and months.
MCP (Model Context Protocol), created by Anthropic, is a standardized translator that lets agents talk to external tools -- Gmail, Notion, Stripe, Calendar -- without custom development for each one.
Skills are markdown-based SOPs stored as .skill files. Once a process is documented in a skill, it never needs to be explained to the agent again -- it executes that process the same way every time.
Skills can be created two ways: feeding a course or reference document to the agent and asking it to generate a skill, or working through a process once manually and asking the agent to package it.
Skills and scheduled tasks can be chained together, enabling fully automated workflows -- like a morning brief that fires at 9:00 a.m., pulls from calendar and inbox, and researches any upcoming guests.
The recommended beginner path is Claude Code or Co-work first. Build out context and skills locally, then migrate to more autonomous harnesses like Open Claw once the workflow is stable.
Context engineering has replaced prompt engineering as the key skill. The better your agents.md file, the simpler your prompts need to be -- and the better your results will be.
Quotable Moments
Quotable moments are auto-generated from the transcript. Speaker attribution and quote accuracy should be verified against the original source before republishing or sharing.
Remy Gaskill
"Chat is question to answer. An agent is goal to result."
Why it works: The clearest one-line definition of the agent paradigm shift -- immediately quotable for social posts, slide decks, or client education.
Remy Gaskill
"It's not about prompt engineering anymore. It's about context engineering -- how well can you load your agent with all the information about your business so that your prompt can be stupidly simple?"
Why it works: Reframes the skill that actually matters now, in direct contrast with mainstream AI discourse still focused on clever prompting.
Remy Gaskill
"I don't even enter these tools anymore. I just sit in Claude Code as one central place."
Why it works: A concrete behavioral claim that makes the productivity vision tangible and personal -- not hype, just a habit change.
Greg Isenberg
"You start to fit a week in a day, and then seven weeks in a week. Stack that up over a year and you're going to be miles ahead of everyone else."
Why it works: Captures the compounding productivity argument in a single sentence without hype jargon -- the kind of line that gets screenshotted.
Concepts & Ideas
Core Framework
Chat vs. Agent: The Fundamental Shift
Chat models operate as a back-and-forth exchange -- you ask, it answers, you do the work. Agents operate on a goal-to-result model where you hand off a task and the agent plans, executes, and delivers without requiring your involvement at each step. This is the central shift from AI as a tool you use to AI as a system that works for you.
The Agent Loop: Observe, Think, Act
Every agent, regardless of which harness you use, runs the same internal loop: it observes its current context and available information, thinks about the best next action, and then acts. It repeats this cycle until the task is complete. Understanding this loop is the foundation of working effectively with any agent platform.
The Four Agent Components
An agent is assembled from four elements: the LLM (the reasoning brain -- Claude, GPT, Gemini, etc.), the loop mechanism, connected tools (what it can interact with externally), and context (what it knows about you and the task). Agent harnesses are simply applications that wire these four things together and keep the loop running.
Practical Principles
agents.md: The Context File
Unlike chat models which store memory in the cloud automatically, agents start each session fresh. The agents.md (or Claude.md, Gemini.md depending on the harness) is a markdown file that gets loaded at the start of every session as a persistent system prompt. It contains your role definition, business context, working preferences, and tool usage guidelines -- essentially, the onboarding document for your AI employee.
memory.md: The Self-Improving Loop
A memory.md file, paired with an instruction in agents.md to update it when corrections are made, gives the agent the ability to learn from interactions and retain preferences across sessions. Over time this file accumulates rules that reduce errors -- things like tone preferences, formatting rules, or client-specific instructions -- compounding into a progressively more capable assistant.
Context Engineering Over Prompt Engineering
The skill that drove results in the early chat era -- crafting long, detailed prompts -- has been superseded. What matters now is how comprehensively your context files describe your business, your preferences, and your processes. When the context is rich, prompts can be extremely simple ("write me a cold email") and still produce excellent results.
MCP: The Universal Connector
Model Context Protocol, created by Anthropic, is a standardized interface that lets any agent speak to external tools without custom integration work for each one. Before MCP, connecting Claude to Gmail or Notion required significant custom development. MCP acts as a universal translator, making tool connections simple and consistent across any harness that supports it.
Skills: SOPs for AI
A skill is a markdown file that documents a repeatable process in enough detail that the agent can execute it consistently every time. Once you build a skill for a process -- writing proposals, analyzing ad libraries, referring clients -- you never need to explain that process again. Skills are stored in a .skill file (or a dot Claude folder) and can be invoked explicitly or automatically when relevant.
Advanced Patterns
Skill Chaining
Skills can call other skills. A morning brief skill might check the calendar, and if a podcast guest is scheduled, automatically invoke a podcast research skill to prepare talking points. This chaining allows complex multi-step workflows to be triggered by a single simple command or scheduled task.
Global vs. Project-Level Resources
Both skills and MCPs can be scoped globally (available across all your agent folders) or at the project level (only available within a specific folder). A truncate style skill might be useful globally, while a client-referral skill should be scoped to the relevant project so it does not clutter context in unrelated workspaces.
The AI Operating System Vision
Remy frames the end state as each person having a personal AI OS -- a folder structure of agents, each managing a different department or life area, all connected to the same underlying tools and context. Rather than opening individual apps, you interact with one central hub (Claude Code, for example) that orchestrates everything. This OS compounds in value as more skills are added and memory accumulates.
Harness Selection Strategy
Different agent harnesses suit different stages of adoption. Claude Code and Co-work are recommended for beginners due to their simple interfaces and clear display of the agent loop. Open Claw offers more autonomous scheduled operation but is harder to configure. The advice: build and refine your skills and context files in a beginner-friendly harness first, then migrate when the workflow is proven.
Implementation
This walkthrough follows Remy Gaskill's exact framework from the course above. Work through it in order. By the end you will have a functional AI executive assistant running on your local machine with memory, tools, and your first skill.
1
Pick a harness and create your project folder
Download Claude Code (easiest for beginners, best loop visibility) or Co-work (clean UI, good for first-timers). Both are free to start. Avoid Open Claw for now -- it is more powerful but harder to configure correctly at the start.
On your computer, create a folder. Name it something clear like executive-assistant. This folder is your agent's workspace. Everything you put in here -- context files, memory, skills -- becomes available to the agent when you point it at this folder.
Why local folders?
Local markdown files are the most durable and portable format for your agent stack. You can move them between harnesses, sync them with Syncthing or Dropbox, and version-control them with Git. Your AI OS lives on your machine, not inside any single app.
2
Create your agents.md context file
This is the most important file in your stack. It is the onboarding document for your agent -- loaded at the start of every session before the agent does anything else. Without it, the agent knows nothing about you.
Open a regular Claude chat (not the agent harness) and paste in this prompt to get started:
Starter prompt -- paste into Claude chat
I want to build an agents.md file to onboard an AI executive assistant. Please interview me with one question at a time to extract the following information: my name and role, what my business does, who my clients or customers are, my working preferences and tone, the tools I use day to day, how I like to communicate externally, and any common tasks I do repeatedly. Once you have gathered all of this, compile it into a clean agents.md file I can save locally.
Answer each question honestly. When Claude produces the file, save it as agents.md (or Claude.md if using Claude Code) inside your executive-assistant folder.
File naming by harness
Claude Code reads Claude.md -- Gemini reads Gemini.md -- Codex and Open Claw read agents.md. Same concept, different filenames. When in doubt, check the harness docs.
Minimal agents.md structure
# Executive Assistant -- [Your Name]
## Role
You are my executive assistant. Your job is to help me manage my day, handle communications, and keep projects moving.
## About Me
[Your name, location, what you do]
## My Business
[What you sell or offer, who you serve]
## Ideal Customer
[Who they are, what they need, how you help]
## Tools I Use
- Gmail -- for all external communications
- Notion -- for project tracking
- Google Calendar -- for scheduling
- Stripe -- for payments
## Working Preferences
- Tone: direct, professional, no fluff
- Sign off emails with: [Your preferred sign-off]
- Never contact clients via Slack -- email only
## Memory and Learning
Before answering any question or starting any task, read memory.md. This contains what you have learned about my preferences over time. When I correct you or share a new preference, update the relevant section in memory.md immediately. Keep memory.md current -- replace outdated information rather than appending duplicates.
3
Add your memory.md file
Create a second file in the same folder called memory.md. It can start empty or with a few section headers. The agent will populate it over time as it learns your preferences.
Starter memory.md
# Memory -- [Your Name]
Last updated: [auto-updated by agent]
## Communication Preferences
[Agent will fill this in over time]
## Formatting Preferences
[Agent will fill this in over time]
## Client Notes
[Agent will fill this in over time]
## Corrections Log
[Agent will note any preferences you have corrected]
The instruction in your agents.md (from Step 2) tells the agent to read this file before every session and update it when you correct it. This is what creates the self-improving loop.
Best practice: keep it under 200 lines
Remy recommends keeping agents.md under 200 lines. If memory.md grows large, add an instruction like "only save substantial corrections -- ignore minor one-off preferences."
4
Connect your tools via MCP
Open your harness app and find the connectors or integrations section. Connect Gmail, Google Calendar, and Notion to start -- these three give you the most immediate productivity gains. Each connection is a simple OAuth sign-in flow.
Once connected, your agent can read from and write to these tools on your behalf. It can summarize your inbox, check your calendar, create Notion pages, or send a draft email -- all from a single natural-language instruction in your agent session.
Security scope
Start with the minimum permissions needed. Read-only access to your calendar and inbox is safer for a first run. Grant write permissions (like sending email) only after you are comfortable reviewing the agent's outputs.
5
Run your first real task
Open the agent harness, point it at your executive-assistant folder, and start a new session. The agent will automatically load your agents.md and memory.md as part of its observe step.
Try a task that uses your connected tools -- for example:
Example task
Summarize my inbox from today, flag anything that needs a reply, and check if I have any meetings tomorrow.
Watch the agent work through the loop. You will see it access Gmail, parse the emails, access Calendar, compile the information, and return a summary. This is the agent loop in action with real tools and your real data.
6
Build your first skill
The next time you complete a repeatable process with the agent, capture it as a skill. A skill is just a markdown file that documents the process in enough detail that the agent can reproduce it without further explanation.
After completing the task, type:
Skill creation command
Use your skill creator skill to create a skill for what we just did. Package up the full process so I can invoke it in future sessions with a single command.
The agent will generate a .skill file and save it to your folder's skills directory. Review it to confirm it captured the process correctly. Good first skills to build: a daily brief, a client follow-up email, a new project setup sequence, or a weekly inbox triage.
Skill creation from a document
You can also create skills from existing reference material -- a course transcript, an SOP document, a style guide. Upload the material and say: "Based on this document, create a skill called [name] that I can invoke to apply this process."
7
Schedule your first automated task
Once you have a skill that runs on a regular basis -- like a daily brief -- set it up as a scheduled task inside your harness. In Co-work and Claude Code, this is now built in to the interface.
Create a new scheduled task and set it to run your daily brief skill every morning at 9:00 a.m. The agent will check your calendar, summarize your inbox, and surface anything that needs attention -- all before you sit down to start working. This is where agents start to feel genuinely autonomous.
8
Expand to additional agent folders
Once your executive assistant is running reliably, replicate the structure for other roles in your business. Create a head-of-marketing folder, a content-team folder, or a folder per client. Each gets its own agents.md scoped to that role, its own memory.md, and skills specific to that function.
This is the full AI OS: a structured set of specialized agents, each operating within a clear scope, all drawing on the same underlying MCP tool connections. At this stage, you are no longer just using AI -- you are running an AI-powered operation.
Agent Builder Master Prompt
The AI prompt on this page is auto-generated from the transcript content and is intended to support further exploration of the topics, concepts, and conclusions discussed. It is provided for informational purposes only. The user is solely responsible for all outcomes resulting from its use.
Master Prompt -- Build Any Agent Using Remy Gaskill's Framework
You are an expert AI agent architect who works exclusively within Remy Gaskill's agent framework as taught in the Greg Isenberg podcast episode "Building AI Agents That Actually Work." Your job is to help me design and build a fully functional agent from scratch, producing all the files I need to drop into my chosen harness and start working immediately.
THE FRAMEWORK YOU WORK WITHIN
Every agent you help me build uses this exact five-layer stack:
Layer 1 -- The Agent Loop
The observe-think-act cycle is the engine of every agent. You understand that any agent platform (Claude Code, Co-work, Codex, Antigravity, Open Claw, Manas) is just a harness that facilitates this loop. When you design an agent, you design for this loop -- giving it the right context to observe, the right goals to think toward, and the right tools to act with.
Layer 2 -- agents.md (or Claude.md / Gemini.md)
This is the onboarding document loaded at the start of every session. It contains: the agent's role definition, business or personal context, working preferences, tool usage guidelines, communication rules, and a memory instruction block. You always produce a complete, specific agents.md -- never a template with placeholders. Keep it under 200 lines unless the complexity of the role genuinely requires more. If context is large, you split it into a context folder and reference it from agents.md.
Layer 3 -- memory.md
A file the agent reads at session start and updates when it learns a preference or receives a correction. You always include a memory instruction block in agents.md that tells the agent to read memory.md before every task and update it when something new is learned. You produce a starter memory.md with section headers ready for the agent to populate.
Layer 4 -- MCP Tool Connections
You know which tools are available via MCP and what each one enables. Core tools: Gmail (read, draft, send), Google Calendar (read, create events), Notion (read, create, update pages and databases), Stripe (create products, payment links, retrieve data), Granola (read meeting notes), Google Drive (read, create files), Slack (read, post messages). When designing an agent, you specify exactly which MCP connections are needed and what the agent will use each one for.
Layer 5 -- Skills (.skill files)
Skills are markdown-based SOPs. Each skill has a clear trigger (when to use it), a step-by-step process the agent follows, and any reference material needed. You produce complete, specific skill files -- not outlines. You know skills can be chained (one skill can invoke another) and scheduled (set to run automatically at a specified time).
HOW TO HELP ME BUILD AN AGENT
When I describe a role or use case, interview me with one question at a time to extract everything you need. Gather:
1. What role this agent plays (executive assistant, head of marketing, client project manager, research analyst, etc.)
2. Business or personal context (what I do, who I serve, what the agent needs to know about my world)
3. The primary tasks this agent will handle day to day
4. Which tools it needs access to (and therefore which MCPs to connect)
5. Communication preferences (tone, sign-off, what channels, what to avoid)
6. Any recurring processes that should become skills from day one
7. Which harness I am using (so you can name the context file correctly)
Once you have enough information, produce:
OUTPUT A -- agents.md
Complete, specific, under 200 lines. Role definition at the top. Business context. Tool usage map. Working preferences. Memory instruction block. If I am using Claude Code, name it Claude.md. If Gemini, Gemini.md. If Codex or Open Claw, agents.md.
OUTPUT B -- memory.md
Starter file with section headers appropriate to this agent's role. Ready for the agent to populate over time.
OUTPUT C -- Skill files
One complete .skill file for each recurring process I have identified. Each skill includes: skill name, trigger phrase, step-by-step process, tools used, and any reference data needed.
OUTPUT D -- MCP connection list
A plain list of which MCP tools to connect in my harness, and one sentence on what the agent uses each one for.
OUTPUT E -- First week prompt sequence
5 to 8 paste-ready prompts I can run in order during my first week to test the agent, build its memory, and establish its first skills.
HARNESS GUIDANCE
When I ask which harness to use, your default recommendation is Claude Code for beginners. If I need scheduled autonomous tasks, you recommend adding Co-work or looking at Open Claw once the workflow is proven in Claude Code first. You know that all harnesses run the same loop and use the same file-based context system -- the only differences are UI and scheduled task support.
WHAT YOU DO NOT DO
You do not produce generic templates with placeholders. Every agents.md, memory.md, and skill file you produce is specific to the role and context I have given you. You do not recommend tools or platforms outside Remy's framework. You do not skip the interview process and guess at context -- if you do not have enough information to write a specific file, you ask for it.
TONE
Direct and practical. Produce files that are ready to use, not ready to customize. If something is unclear, ask one specific question before proceeding.
[Tell me what role or department you want to build an agent for, and I will start the interview.]
Skill Builder Master Prompt
The AI prompt on this page is auto-generated from the transcript content and is intended to support further exploration of the topics, concepts, and conclusions discussed. It is provided for informational purposes only. The user is solely responsible for all outcomes resulting from its use.
Master Prompt -- Build Any Skill File Using Remy Gaskill's Framework
You are an expert AI skill architect who works exclusively within Remy Gaskill's skill framework as taught in the Greg Isenberg podcast episode "Building AI Agents That Actually Work." Your job is to help me design and document any repeatable process as a .skill file that my agent can invoke consistently, without further explanation, every single time.
WHAT A SKILL IS
A skill is a markdown-based Standard Operating Procedure for an AI agent. Once a skill exists, the agent can execute that process the same way every time -- no re-explaining, no going back and forth, no variation. Skills are the compounding layer of an AI operating system: the more skills you build, the more of your workflow runs on autopilot.
Skills live in a .skill file (or inside a dot Claude folder depending on your harness). They can be:
- Scoped to a specific project folder (used only in that agent context)
- Scoped globally (available across all your agent folders and projects)
- Chained with other skills (one skill can trigger another as part of a larger workflow)
- Scheduled (set to run automatically at a specified time via your harness)
THE TWO CREATION METHODS REMY USES
Method 1 -- From a reference document or course
You have existing material (a course transcript, an SOP document, a style guide, a process manual) and you want to turn it into a skill the agent can follow. You upload or paste the material, describe the output you want, and the agent packages it as a skill.
Method 2 -- From a live process
You work through a task manually with your agent once. At the end, you say: "Use your skill creator skill to create a skill for what we just did." The agent documents the steps it took, the tools it used, and the decisions it made, and packages that into a reusable skill file.
You support both methods. When I give you reference material, you extract the process and structure it as a skill. When I describe a process I have done manually, you ask clarifying questions and then document it as a skill.
SKILL FILE STRUCTURE
Every skill you produce follows this structure:
---
skill: [skill-name]
trigger: [the phrase or condition that invokes this skill]
scope: [project | global]
tools: [list of MCP tools this skill uses, if any]
chains: [list of other skills this skill invokes, if any]
schedule: [cron expression or plain-language schedule, if applicable]
---
## Purpose
One sentence. What does this skill do and why does it exist.
## When to Use
The exact condition or trigger phrase that causes the agent to invoke this skill. Be specific -- ambiguous triggers cause the wrong skill to fire.
## Reference Material
Any data, templates, formulas, or examples the agent needs to execute this skill. Paste them directly into the skill file so the agent does not need to look elsewhere.
## Process
Numbered steps. Each step is specific enough that the agent can execute it without further instruction. Where a step uses a tool (Gmail, Notion, Stripe, etc.), name the tool and describe the exact action. Where a step involves judgment (e.g. "assess tone"), specify the criteria for that judgment.
## Output
What the agent produces at the end of this skill. File format, destination, or action (e.g. "a drafted Gmail message in the Drafts folder" or "a new Notion page in the Projects database").
## Quality Check
2 to 4 criteria the agent should verify before considering the skill complete.
HOW TO HELP ME BUILD A SKILL
When I describe a process or share reference material, interview me with one question at a time to gather everything you need. For each skill, you need to know:
1. What process this skill automates (describe it step by step as you currently do it manually)
2. What triggers this skill (the exact phrase I will say, or the condition that should fire it automatically)
3. Whether it should be project-scoped or global
4. Which tools it needs (Gmail, Notion, Stripe, Calendar, Granola, Google Drive, Slack, web search, etc.)
5. Whether it chains with any other skills
6. Whether it runs on a schedule (and if so, when)
7. What the final output looks like (what I receive or what gets created)
8. Any reference material, templates, or examples the agent needs to execute it well
Once you have enough, produce the complete .skill file. Do not produce outlines or templates -- produce the finished file, ready to drop into my skills folder.
SKILL QUALITY STANDARDS
A well-built skill meets all of these:
- The trigger is unambiguous -- the agent will only fire this skill when you actually want it
- Every step is specific enough to execute without clarification
- All reference material is embedded in the file, not referenced externally
- Tool actions are named precisely (e.g. "use Gmail MCP to create a draft addressed to [contact from memory.md]" not just "send an email")
- The output is clearly defined so the agent knows when it is done
- The quality check gives the agent a way to verify success before closing the loop
COMMON SKILL TYPES TO BUILD FIRST
If I am just getting started and not sure what to build, you suggest these high-value first skills based on Remy's framework:
- Daily brief -- checks calendar, summarizes inbox, surfaces action items, plans the day
- Client follow-up -- pulls meeting notes from Granola, drafts a follow-up email with next steps, creates a Notion project if one does not exist
- Proposal -- produces a formatted proposal document from a client brief, with pricing at the bottom
- Weekly research -- scrapes Twitter and Reddit (or specified sources) for developments in a topic area, produces a formatted briefing
- New lead intake -- when a new lead comes in, logs it to Notion, drafts a first-contact email, and adds a calendar reminder to follow up in 48 hours
- Content brief -- takes a topic, researches it via web search, and produces a structured content brief with angle, hook, key points, and call to action
CHAINING EXAMPLE
A morning brief skill might invoke: a calendar check sub-skill, a podcast guest research skill (if a recording is on the calendar), and an inbox triage skill. You know how to design these chains and document them clearly in the chains field and in the process steps.
SCOPE GUIDANCE
Global skills: anything you use across multiple agent folders (truncate, format, summarize, tone-shift, research)
Project skills: anything specific to one client, one role, or one workflow (client-name follow-up, specific proposal format, platform-specific ad analysis)
WHAT YOU DO NOT DO
You do not produce generic skill templates with placeholders. Every skill file you produce is specific to the process I described, the tools I use, and the output I need. You do not skip the interview if you are missing critical information -- you ask one specific question at a time until you have what you need to produce a finished, usable file.
TONE
Direct, practical, zero fluff. Produce skill files that work on the first invocation. If a process has ambiguity in it, surface that ambiguity in your questions before writing the file -- not after.
[Describe the process you want to turn into a skill, or paste in the reference material you want packaged as a skill, and I will get started.]
Tools & Platforms
Every tool, platform, and service mentioned or recommended in this episode. Organized by category.
Anthropic's agentic coding and task environment. Remy's primary recommendation for beginners -- displays the agent loop clearly so you can see exactly what the agent is doing at each step. Uses a Claude.md file for context. Supports MCP tool connections and scheduled tasks.
OpenAI's agent harness. Uses an agents.md file for context. Demonstrated alongside Claude Code in the live build demo. Operates on the same observe-think-act loop as all other harnesses.
Agent harness used in the live demo to build a portfolio site. Produced the cleanest visual result in the side-by-side comparison. Works off the same local folder and context file model.
The harness Remy uses to demonstrate the executive assistant build in this episode. Praised for its simple, beginner-friendly UI and clear connector/integration management. Supports scheduled tasks and MCP connections. Good starting point alongside Claude Code.
The most autonomous harness discussed -- supports cron-based scheduled tasks and more hands-off operation. Remy's recommendation is to build and test your agent stack in Claude Code first, then migrate to Open Claw once the workflow is stable. Uses a dot open claw folder and the same markdown context file pattern.
Mentioned as a harness with a simple connector interface for MCP tools. Remy has Perplexity connected via MCP for research tasks, as demonstrated when the agent autonomously researched Greg Isenberg during the portfolio site build.
Connected via MCP for reading, drafting, and sending email. Core to the executive assistant demo -- the agent reads incoming mail, drafts follow-ups, and sends on command. Remy's stated preference is to keep all client communications on email, enforced via the agents.md file.
Remy's project management tool, connected via MCP. In the demo, the agent creates a new project in Notion automatically after a prospect call -- pulling meeting notes, drafting the proposal email, creating the Stripe link, and setting up the Notion project in one command.
Connected via MCP for creating products and payment links. Demonstrated live -- after a prospect meeting, the agent creates a Stripe payment link as part of the proposal workflow without the user ever opening Stripe.
AI meeting notes tool. Connected via MCP so the agent can read meeting notes and pull context from recent calls. Used in the demo to extract the prospect meeting summary before drafting the follow-up email.
Mentioned as an MCP connection for file access. Remy references using it in his ads analysis skill to access creative assets from Dropbox or Drive as part of the process.
Connected via MCP as a research tool. Demonstrated when the agent autonomously launched a sub-agent to research Greg Isenberg before building his portfolio site -- without Remy asking it to do so explicitly.
Remy's preferred voice-to-text tool. Hold a button, speak, and it transcribes neatly into text -- used to issue commands to the agent without typing. Built by the team at Every.io. Greg noted it looks well-designed.
Referenced as a popular way to extend the context file system. Remy notes that many people point their agents.md at their Obsidian vault as a second brain -- instructing the agent to check the vault before answering questions or starting tasks.
The LLM behind Claude Code and the primary model discussed throughout the episode. Anthropic also created MCP (Model Context Protocol), the connector standard that makes tool integration straightforward across all harnesses.
Mentioned as another LLM option. Uses a Gemini.md context file when running in an agent harness. In the side-by-side demo, Greg indicated he preferred Gemini's portfolio site output.
Created by Anthropic. The standardized protocol that lets any agent speak to external tools without custom integration work per tool. Acts as a universal translator between your agent and apps like Gmail, Notion, Stripe, and Slack. All major harnesses support it. This is the infrastructure layer that makes the connected agent OS possible.
Not mentioned in this episode but referenced in the broader Remy framework -- a peer-to-peer file sync tool that keeps local agent folders in sync across machines without a cloud intermediary. Relevant for anyone running their AI OS across a desktop and laptop.
Referenced in the context of Remy's ads analysis skill, where the agent accesses ad creative assets stored in a Dropbox folder as part of the analysis process.
Guest and framework author. AI productivity practitioner who runs multiple business lines using the AI OS system described in this episode. Follow for ongoing skill and agent framework updates.
Referenced as a previous podcast guest who coined the idea of a "100X employee" built on a pre-existing AI OS -- a vision Remy echoes throughout this episode.
Transcript
This transcript was auto-generated and may contain errors in speaker attribution, transcription accuracy, or formatting. Confirm accuracy and completeness against the original source before referencing or republishing.
Greg Isenberg: I think AI is confusing. There, I said it. I think there's a lot of terms -- skills, MCPs, agent harnesses -- that are difficult concepts to understand. So I had my friend Remy come on the podcast and explain it in the most simple terms possible. In this free course on how to master AI agents, he breaks down exactly what each piece is, how they connect together, and the simplest ways beginners could start using them today. Enjoy the episode.
Greg: I begged them to come on. Remy Gaskill's on the pod. You've structured your company where you basically have these folders and dot MD files that run your company. And what I want to do today is I want you to teach people in a beginner-friendly fashion -- this is only for beginners -- how they could do the same thing. How they can set up their own executive assistant, head of marketing, chief financial officer. By the end of this episode, Remy, do you think you can do that?
Remy Gaskill: 100% Greg. We're going to go through all the concepts that make up an AI agent. And by the end of this video, you will know exactly how you can build up agents to run complete departments of your life and your company within any agent platform you choose, whether it's Claude Code, Codex, Open Claw, Manas -- all of them.
Remy: The key thing to understand is chat models versus agents. A chat model is question to answer. An agent is goal to result -- moving from you asking and AI replying to you giving the agent a task, it planning out the task, executing, and then delivering you a result.
Greg: Crystal clear. Chat is kind of like ping pong -- back and forth. An agent is you giving it a goal and hoping that over time it gets better and closer to that goal.
Remy: Exactly. Inside the agent, we have the agent loop -- three steps: observe, think, and act. The agent receives your prompt, checks its workspace, thinks about the best next action, acts, feeds the result back in, and keeps going through this loop until the task is complete. The loop continues based on parameters you set in your prompt -- like "compile 10 sources and create a PowerPoint report."
Remy: An agent is made up of four components: the LLM (the brain -- Claude, GPT, Gemini), the loop (continuous execution until done), connected tools, and context. A platform that facilitates this process is known as an agent harness. All popular AI agent platforms are just agent harnesses. Understanding the loop means you can operate any of them -- like knowing how to drive means you can get into any car.
[00:15:00]
Remy: My workspace looks like this: a big folder for each company or client, folders underneath with all my heads of departments, and within those, skills and MCPs. We're going to be focusing today on building out an executive assistant to take care of your manual day-to-day tasks and free up at least one to two hours per day.
Remy: The way I like to think about building agents is onboarding them like a real employee. If you took on a real executive assistant, you couldn't expect them to just get a task without you first explaining your business, your clients, what you do, the tools. So that's the first step.
[00:20:00]
Remy: We haven't populated an agents.md file. An agents.md file is just like a system prompt -- it gives the agent context that's always on. You put in there its role, context about you, your preferences. Every new session, before it answers your query or task, it loads in all this context as part of the observe step in the loop. When working within Claude Code, it's called a Claude.md. When working within Gemini, it's a Gemini.md. But it's all the same concept.
Remy: One of the other big shifts when moving from chat to agents: prompt engineering used to be the big thing. Now it's all about context engineering. It's about how well you can load up your agent with all the information about your business so that your prompt can be stupidly simple -- like "write me a cold email" -- and you still get an amazing result.
[00:25:00]
Remy: A new problem: the agent doesn't really remember intricate details or your preferences across sessions unless you're manually updating the Claude.md file. So what I like to do is add a memory.md file and an instruction in agents.md to read it and update it whenever I correct the agent. This creates a self-improving loop that compounds over weeks and months.
[00:30:00]
Remy: Now we need to connect our tools because by default most agent harnesses just have web search baked in. If you want to link it up to Gmail, calendar, and everything else -- which is where the real productivity gains are made -- you need to do so via MCP. Anthropic built MCP to sit as a translator between your tools so that Claude can just speak English, your tools can speak their languages, and MCP speaks every language and translates between them. A really easy standardized way to connect tools up.
[00:35:00]
Remy: I've connected Gmail, Google Calendar, Granola for meeting notes, Stripe for payments, Notion for project management. And I don't even enter these tools anymore. I just sit in Claude Code as one central place. For example: "Review my meeting notes with Maltoshi from today, draft the email sending the proposal and creating the Stripe payment link, and then go into Notion and set up the project."
Remy: And where this starts to compound even more is when you start to build out skills for each of your processes. Every time I do a process manually and I know I'm going to do it again, I turn it into a skill. If you automate three to five tiny manual processes each week with skills, you eventually end up automating your entire life with these agents.
[00:40:00]
Remy: The easiest way to think about skills: SOPs for AI. Standard Operating Procedures. Once you explain something once, you never have to explain it ever again. A skill packages up a process into a .skill file -- a markdown file that explains the exact process you went through. Every time you invoke that skill, it executes the process the same way every single time.
Remy: Two ways to create skills. One: you can take a course or reference document, transcribe it, put it into Claude, and say "create a skill from this." Two: go through a process manually once with Claude, and then say "create a skill for what we just did." It packages up the whole process you went through.
[00:50:00]
Remy: You can chain skills together. A morning brief skill might check the calendar and, if there is a podcast guest scheduled, automatically invoke a podcast research skill. You can also schedule skills -- set a morning brief to run every day at 9:00 a.m. in Co-work or Claude Code. That is where agents start to feel genuinely autonomous.
[00:55:00]
Remy: For the beginner: Open Claw is probably the hardest to learn. Claude Code is probably the easiest. Build and refine everything in Claude Code first. Once you're happy with how it's functioning, then look to migrate to Open Claw where it has that more autonomous nature. Global versus project-level: skills, MCPs, and Claude.md files can all be scoped globally or per project. Use global for things you need in every session, project-level for context specific to one client or role.
Remy: That's pretty much the entire agents crash course. It's just that loop running in the backend to complete your task, connecting in your tools, your context, and the LLM all in one place. Go into Claude or your favorite chat model and get it to help you build out those context files through an interview-style process. Connect all the tools you need and then start building out the skills through daily use.
Greg: Remy, thank you so much. I appreciate you coming on and dropping some knowledge. Thank you, man.
Remy: Thank you so much for having me on, Greg. It's been a blast.
Exploration Prompt
The AI prompt on this page is auto-generated from the transcript content and is intended to support further exploration of the topics, concepts, and conclusions discussed. It is provided for informational purposes only. The user is solely responsible for all outcomes resulting from its use.
Exploration Prompt -- Building AI Agents That Actually Work
You are an expert AI agent architect and implementation coach, trained on the full framework presented by Remy Gaskill in the Greg Isenberg podcast episode "Building AI Agents That Actually Work."
CONTEXT AND FRAMEWORK
This conversation is grounded in a specific, practical framework for building AI agents that actually function in a real business context -- not theoretical concepts, but a working operational system that Remy runs to manage multiple companies and clients. The core framework has five layers:
1. The Agent Loop -- the observe-think-act cycle that every agent runs internally. Understanding this loop is the foundation. Chat models are ping-pong (one prompt, one reply). Agents are goal-to-result: you hand off a task, the agent plans and executes it autonomously until completion.
2. Context Files (agents.md / Claude.md) -- the onboarding document for your agent. Loaded at the start of every session. Contains your role definition, business context, working preferences, tool usage, and communication guidelines. Without this, the agent knows nothing about you. Context engineering -- building rich, accurate context files -- has replaced prompt engineering as the key skill.
3. Memory (memory.md) -- a file the agent updates when it learns your preferences or receives corrections. Paired with an instruction in agents.md to read and update it, this creates a self-improving loop that compounds over weeks and months. Best practice: keep agents.md under 200 lines; use a separate context folder for deeper reference material.
4. Tools (MCP) -- Anthropic's Model Context Protocol is a universal translator that connects agents to external tools (Gmail, Notion, Calendar, Stripe, Slack) without custom development per tool. Once connected, the agent can read from and write to these tools as part of executing any task.
5. Skills (.skill files) -- markdown-based Standard Operating Procedures for AI. Once a process is documented in a skill file, the agent can execute it consistently without re-explanation. Skills compound: the more you build, the more of your workflow is automated. They can be chained together and scheduled for fully autonomous operation.
KEY PRINCIPLES
- An agent harness (Claude Code, Co-work, Codex, Antigravity, Open Claw) is just an application that runs the agent loop. Learn the loop, and you can operate any harness.
- Onboard your agent like a real employee. You would not hand a new assistant a task on day one without first explaining your business, clients, tools, and preferences.
- Memory and context are intentional. Unlike chat models that store memory automatically in the cloud, agent memory must be explicitly set up and controlled -- which is a feature, not a limitation.
- Skills are built two ways: from existing reference material (courses, SOPs, style guides) or from a live process you have completed once and want to repeat.
- Build in Claude Code or Co-work first. Migrate to Open Claw for autonomous scheduled operation only after the workflow is proven and stable.
- Global vs. project-level scope: some skills, MCPs, and context files belong globally (used in every workspace); others should be scoped to a specific project or client folder.
WHAT THIS IS NOT
This is not about using ChatGPT in a chat window more cleverly. It is not about writing better prompts. It is not about subscribing to an AI tool and hoping for the best. This framework requires you to build a local folder structure, write context files, connect real tools, and progressively build skills for your specific processes. It is an investment that compounds -- and the returns are proportional to the effort you put into the context and skills layers.
HOW TO USE THIS CHAT
1. Build your context file -- Ask me to interview you with questions to help you build an agents.md or Claude.md file for a specific role or use case (executive assistant, head of marketing, client folder, etc.).
2. Design a skill -- Describe a process you do repeatedly. I will help you document it as a .skill file that your agent can invoke consistently.
3. Plan your tool connections -- Tell me what tools you use and what tasks you want to automate. I will map out which MCP connections you need and what your agent will be able to do once they are set up.
4. Troubleshoot your agent -- Describe what is going wrong with your current agent setup (wrong outputs, forgetting preferences, ignoring context). I will help you identify whether the issue is in the context file, the memory setup, the skill design, or the prompt.
5. Chain skills into a workflow -- Describe a recurring daily or weekly workflow. I will help you design a chain of skills and scheduled tasks that automates it end to end.
6. Plan your AI OS -- Tell me about all the roles in your business or life. I will help you map out the folder structure, context files, and skills you need to build a full AI operating system.
TONE
Direct, practical, and specific. No hype. No vague AI talk. Ground every answer in the actual mechanics of files, folders, loops, and tool connections. If I need more information to give a useful answer, ask for it.
[Drop your question here -- or tell me what role or workflow you want to build an agent for, and I will walk you through it step by step.]