Greg Isenberg -- Master AI Agents Crash Course

Concepts and Ideas

Core Framework -- How Agents Work

Chat vs. Agent

Chat models are reactive -- you prompt, they answer, you do the work. Agents are autonomous -- you give a goal, the agent plans and executes until it delivers a result. This shift from ping-pong to goal-driven loops changes what you need to do and what the AI handles.

The Agent Loop (Observe -- Think -- Act)

Every agent operates an internal loop: observe (read the current state, files, tools, prior outputs), think (decide the next step), act (execute that step). This loop repeats as many times as needed until the task parameters are satisfied. Understanding this loop explains why agents can handle multi-step tasks without prompting at each stage.

The Four Agent Components

An agent is composed of: the LLM (the reasoning brain), the loop (the execution engine), tools (external connections), and context (everything the agent knows). All popular agent platforms are simply harnesses that facilitate these four elements. The harness choice is a preference -- the underlying components are universal.

Agent Harnesses

Platforms like Claude Code, Codex, Anti-Gravity, Cowork, Manus, and OpenClaw are all agent harnesses -- applications that facilitate the agent loop. They differ in UI, feature maturity, security model, and autonomy level, but the core mechanics are the same. Learning the concepts in one harness transfers to all others.

Context and Memory

agents.md / claude.md -- The System Prompt File

This markdown file is the always-on context your agent loads before every session. It contains role definition, business context, working preferences, and tool information. It is the equivalent of a custom GPT's system prompt -- except it lives locally on your machine and applies within any harness you work in. Without it, the agent has no idea who you are or what it should do.

memory.md -- The Persistent Learning File

Agents do not automatically remember corrections or preferences across sessions unless you design a memory system. By instructing the agent in agents.md to write learned preferences and corrections into memory.md, you create a compounding knowledge base. Every correction compounds -- the agent gets more accurate over weeks and months without manual intervention.

Context Engineering (Not Prompt Engineering)

The dominant skill for working with agents is not writing clever prompts -- it is loading rich, accurate context into the agent's operating environment. When context is thorough, prompts can be minimal ("write me a cold email") and results are still high-quality because the agent knows everything it needs to complete the task. Context quality is the primary variable in agent output quality.

Tools and Integration

MCP -- Model Context Protocol

Built by Anthropic, MCPs are a standardized translation layer between your agent and external tools. Before MCPs, connecting an LLM to a tool required bespoke development for each integration. MCPs work as universal translators -- the agent speaks to the MCP, the MCP translates to the tool's native language and back. This enables practical integrations with Gmail, Google Calendar, Notion, Stripe, Granola, and hundreds of other services with minimal setup.

Global vs. Project-Level Configuration

Context files, MCPs, and skills can be set at a global level (applying to every project) or a project level (scoped to a specific folder or agent). A truncation skill used everywhere belongs at global level. A Sebastian-referral skill specific to one business relationship belongs at project level. Managing this scope deliberately keeps agent context focused and prevents irrelevant information from cluttering the working environment.

Skills and Automation

Skills -- SOPs for AI

A skill is a markdown (.skill) file encoding a completed workflow or process. Once built, invoking the skill name tells the agent exactly what to do, in what order, with what parameters -- without re-explanation. Skills are the primary mechanism for turning one-off agent assistance into repeatable, reliable automation. Three to five new skills per week compounds into a substantially automated work life within months.

Skill Chaining

Skills can invoke other skills. A morning brief skill can detect upcoming meetings and automatically trigger a meeting prep skill and a podcast research skill. Chained skills allow complex, multi-step workflows to run from a single command -- or on a scheduled task -- without any human intervention at each step.

Scheduled Tasks

Most agent harnesses now support scheduled task execution (cron-style). Skills can be set to run at specific times -- a daily morning brief at 9am, a weekly research scrape every Thursday, a marketplace monitor every three hours. Scheduled tasks transform the agent from a responsive tool into a background worker that runs whether you are at the keyboard or not.

The AIOS (AI Operating System)

Remi's framing for where this all leads: a personal folder-based operating system that compounds over time. It is a structured directory of context files, memory files, tool connections, and skills organized by department or role (executive assistant, content team, ads manager, etc). Each week, new skills reduce manual effort. Over months and years, the system handles an increasing share of the operational work in a business or personal workflow.

Implementation Steps

Implementation steps are auto-generated from the transcript content and are provided for informational purposes only. They do not constitute professional advice of any kind. Always consult a qualified professional before acting on any information presented here.

Pick a harness and learn it first

Choose one agent platform to start with -- Cowork or Claude Code are the most beginner-accessible. Resist the temptation to try multiple harnesses at once. The concepts (agents.md, memory.md, MCPs, skills) are identical across platforms, so learning one well means the others come easily. Get comfortable with the folder-based workflow and the agent loop before branching out.

Build your agents.md foundation file

Open a chat session and ask Claude to interview you and build an agents.md file from your answers. Cover your role, what your business does, what clients you serve, which tools you use, and how you prefer to work. This file becomes the permanent onboarding document your agent loads before every task. Treat it like an employee handbook for your AI. Keep it under 200 lines and update it as your work evolves.

Add memory.md with update instructions

Create an empty memory.md file in the same folder. At the top of your agents.md, add an instruction telling the agent to read memory.md at the start of every session and to write any learned preferences, corrections, or important updates to it. This creates the self-improving loop. The first few sessions will feel similar to before -- the compounding becomes noticeable over weeks as corrections accumulate.

Connect your most-used tools via MCP

Inside your chosen harness, go to connectors or settings and connect the tools you actually use -- email, calendar, project management, payments, meeting notes. Start with three to five that you currently copy-paste context from manually. Once connected, the agent can read, write, and act across these tools from a single prompt without you switching applications. This is where the time compression becomes concrete and measurable.

Identify your first three skill candidates

Look at the last two weeks of work. What tasks did you do more than once? What processes involved multiple steps, multiple tools, or multiple rounds of back-and-forth correction with the AI? Those are skill candidates. Start with the smallest one -- something that takes 10 to 20 minutes manually -- and plan to build it as a skill first. Tiny skills compound faster than ambitious ones because they get used more frequently.

Build your first skill by doing the task once

Complete the target process manually inside your agent session -- prompt by prompt, correction by correction -- until you have a result you are happy with. At the end, tell the agent to use the skill-creator skill to package everything you just did into a named skill file. Review the skill file it produces, add any missing details, and save it to your skills folder. From that point forward, invoking the skill name replicates the entire process without explanation.

Set up at least one scheduled task

Pick a recurring workflow that benefits from running without a manual trigger -- a morning brief, a weekly research pull, a marketplace monitor, a social media scan. Write the skill for it, then create a scheduled task in your harness that invokes that skill at the desired time. Once running, this is a zero-effort background process. The discipline of building scheduled tasks moves agents from assistant to autonomous operator.

Organize by department or role folder

As you build more skills and context, organize them into separate folders by function -- executive-assistant, content-team, research, finance, etc. Each folder gets its own agents.md tailored to that role's context and responsibilities. This mirrors how you would onboard a real department head: different scope, different instructions, different tools. An overarching orchestrator agents.md at the top level can delegate across departments when needed.

Build three to five new skills per week

Adopt the habit of ending any new multi-step process with "create a skill for what we just did." At three to five skills per week, within a month you have a meaningful library of automated processes. Within six months, the majority of your repetitive work is handled by skills that run reliably and consistently. The compounding is exponential: more skills mean more chaining possibilities, which create more automation without additional effort.

Migrate to a more autonomous harness when ready

Once your skills are validated and your context files are solid in Claude Code or Cowork, you can migrate the stack to a more autonomous harness like OpenClaw for specific departments. OpenClaw has a steeper learning curve but supports more autonomous operation. The strategy is to prove and polish everything in an accessible harness first, then hand off the stable workflows to an environment where they can run more independently with less oversight.

Full Transcript

This transcript was auto-generated and may contain errors in speaker attribution, transcription accuracy, or formatting. Long transcripts may be truncated due to processing limits. Confirm accuracy and completeness against the original source before referencing or republishing.

[00:00]

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 Remi 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 Isenberg

I beg them to come on. Remi Gascal on the pod. You've structured your company where you basically have these folders and .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. I want you to tell us the concepts behind all this. By the end of this episode, Remi, do you think you can do that?

Remi Gascal

100%, Greg. We're going to go through all the concepts that make up an AI agent. 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, OpenClaw, Manus -- all of them.

Remi Gascal

One of the reasons I really wanted to make this episode is because the AI landscape is moving into stage two -- from chat to agents. Most people are getting left behind right now, just using the chat models. The founders and employees that are utilizing agents are nowhere to rely 10 to 20 times more productive in their day. Stack that up over days, weeks, years, and you're going to be miles ahead of the competition. So the key thing to understand here is chat models versus agents.

Remi Gascal

The way I think of it is: a chat model is question to answer. An agent is goal to result. Moving from question-and-answer ping-pong to giving the agent a task, it planning out the task, executing, and then delivering you a result.

Greg Isenberg

Crystal clear. The way I think of it is chat is kind of like ping pong -- back and forth. An agent, you're giving it a goal and you're hoping that over time it gets better and closer to that goal.

Remi Gascal

Exactly. And inside the agent step we have what's called the agent loop: observe, think, and act. You give it a task -- say, build me a minimalist portfolio site for Greg Isenberg -- and it starts by checking if there are any existing files in the workspace it can work with. It thinks about what to do next, then acts, and keeps going through the loop until the task is complete. How it concludes the task is complete is based on the parameters you set in your prompt.

Remi Gascal

The agent itself is made up of four components: the LLM -- the brain -- think Claude Opus 4.6 or GPT 5.4 or Gemini 3. The loop, which means it keeps going until the task is done. It connects in all your tools. And it connects in all the context. A platform that facilitates this loop is known as an agent harness. Claude Code, Codex, Manus, Anti-Gravity -- they're all just agent harnesses. Applications where the loop is facilitated.

[10:07]

Remi Gascal

A nice analogy: today we're going to learn to drive. We're going to learn about how to steer the car -- pedals, brakes, accelerator. Once you know how to drive, you can jump in any car, whether it's an old Toyota or a Range Rover. The agent harnesses are like different cars. Some will have better features -- seat warmers, cruise control -- but once you know how to drive, you can jump in any of them and use them.

[15:01]

Remi Gascal

My workspace looks something like this: a big folder for each company or client I'm working in, then folders underneath with all my heads of departments. Within those heads of departments I have skills, the MCPs, and context. We're going to focus today on building out the executive assistant to take care of just your manual day-to-day tasks and free up at least one to two hours extra per day.

Remi Gascal

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 come into the office and you'd give them a task without explaining your business first -- your clients, what you do, the tools. So that's the first step: give your agent context.

[20:04]

Remi Gascal

An agents.md file is just like a system prompt. You put in there the agent's role, context about you, your preferences for working. Every new session, before it answers your query or task, it loads in all this context to its brain as part of the observe step in the loop. When you're working in Claude Code it's called claude.md. In Gemini it's gemini.md. In Codex or OpenClaw it's agents.md. Same concept throughout.

Remi Gascal

One of the big shifts moving from chat to agents is prompt engineering used to be the big thing. Now it's all about context engineering. How well can you load up your agent with all the information about your business so that your prompts can be stupidly simple -- like "write me a cold email" -- and you still get an amazing result.

[25:02]

Remi Gascal

Agents have a new problem: they don't really remember intricate details or your preferences across sessions, unless you're manually going and updating that claude.md file. So I add to my agents.md file an instruction: read all files in context, read memory.md -- this is what you've learned over time -- and when I correct you or you learn something new, update the relevant section in memory.md. Keep memory.md current when something changes -- update in place and replace outdated info.

Greg Isenberg

Is a skill like a memory file? Like what's the difference between a memory.md and a skill -- is it just a memory.md for a particular job to be done?

Remi Gascal

That's pretty much exactly it. And all of these agent harnesses now have skills as a feature. A skill is basically like your memory.md which explains the exact process -- except packaged as a reusable file. It packages up that process into a .skill file. In that file it's basically just a markdown file that explains the exact process you went through. So every time you need that process done, the skill knows exactly what to do, delivering the same result every time.

[30:05]

Remi Gascal

A best practice for those claude.md files is to keep it around no more than 200 lines. I could imagine if you started to build this up over years you'd eventually hit a point where all the rules are stepping on each other's toes. You can update your claude.md to say "only save substantial corrections" for more control over what it's saving.

Remi Gascal

By default, most of these agent harnesses just have web search baked in. But if you want to actually start linking it up to your tools like Gmail, Calendar, and everything else -- which is where the real productivity gains are made -- you need to do so via what's called MCP. Before MCPs, your agent, in order to speak to tools, had to kind of learn their language -- Claude speaks English, Notion speaks Spanish, Gmail French, your browser Japanese, Slack Chinese. MCPs sit as a translator in between your tools. Claude still just speaks English and your tools speak their languages -- and MCP translates in both directions. That's what Anthropic built.

[35:00]

Remi Gascal

The way I see the future of this going is everyone's going to have their own AIOS -- an AI operating system -- that compounds over time. Adding tools and then skills. Everyone will have personal agents and agents to manage each department of their company. I don't even enter Gmail, Google Drive, Calendar, Granola, Stripe, Notion anymore. I just sit in Claude Code as one central place.

[40:04]

Remi Gascal

The easiest way to think about skills is SOPs for AI -- standard operating procedures for AI. Once you explain something once you never have to explain it ever again. Without skills, every time you do a proposal you're going back and forth 15 to 30 minutes to get it right. Next week you need another proposal -- unless you found the same session, it's completely forgotten all those preferences. Skills package that process so you get the same result every single time.

Remi Gascal

There are two ways I find useful to create skills. One: you have an idea of a skill off the bat -- like viral hooks. I had a course on viral hooks, transcribed it, put it into Claude, and used the skill-creator skill built into Claude Code and all major harnesses. I said "take this course and create a viral hook skill." It created it like that. Second way: go through a process manually once with Claude, then once you've done the task, say "create a skill for what we just did" -- and it packages up the process you went through.

[45:02]

Remi Gascal

You can chain skills together. A meeting prep skill, a podcast research skill, a morning brief skill -- and in the morning brief skill you can say "if there are any meetings or podcasts in my day, use the podcast research skill to research the guest." You can like chain them together and build really cool workflows.

Greg Isenberg

And you can have it send you an email.

Remi Gascal

Exactly. And these harnesses are starting to add more autonomy -- scheduled tasks. In Cowork or Claude Code now you can set a new task and say "run my morning briefing skill every morning at 9am" and now it's like an automated workflow running every morning. I'm buying a new car right now -- particular color, unique feature set -- and I have it scraping all the different car marketplaces every three hours, notifying me when something comes up. Saves me an hour a day of religious refreshing across Carmax, Cars.com, Autotrader.

[50:02]

Remi Gascal

I built out an agent to manage Meta ads in OpenClaw and it went pretty viral. It runs off an agents.md file the same way but in a .openclawfolder instead of a .claude folder. Same concept -- markdown context files, connecting your tools, creating skills. I planned it out with Claude: "I want this OpenClaw to manage my Meta ads -- help me write the agents.md." Created skills for ad creative, copywriting, maybe 15 different skills. Combined scheduled tasks with skills and context and gave it all the tools it needed. Same process as building the executive assistant.

[55:02]

Remi Gascal

For the beginner: I would say OpenClaw is probably the hardest to learn and set up. Cowork is probably the easiest. Get comfortable using Claude Code or one of these other harnesses before playing around with OpenClaw. And have all the processes built out in Claude Code first. Over two weeks, build out a bunch of skills -- daily brief, meeting prep, the Sebastian refer skill -- and once I'm happy with how it all functions in Claude Code, then look to migrate that into OpenClaw where it has that more autonomous nature.

Remi Gascal

Global versus project level -- just a quick note: skills, context files, MCPs can all be set at a global level (apply to every project) or a project level (scoped to one folder). A truncate skill you use everywhere belongs at global. The Sebastian referral skill for one client relationship belongs at project level. Managing scope deliberately keeps the agent focused and prevents irrelevant context from cluttering the working environment. And that's pretty much the entire agent crash course. The loop running in the backend, connecting your tools, your context, the LLM -- all in one place. Go into Claude or your favorite chat model, get it to help you build out those context files through an interview process, connect all the tools you need, then start building out skills through daily use. Pretty soon you'll have powerful agents built for every department of your business.

Greg Isenberg

Remi, thank you so much. Good links in the show notes and description where you can go follow him, get to know him a little bit better. I appreciate you coming on and dropping some sauce.

Remi Gascal

Thank you so much for having me on, Greg. It's been a blast.

AI 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 -- AI Agents Crash Course (Isenberg / Gascal)

You are an expert guide on building and operating AI agents using the folder-based, markdown-driven framework taught by Remi Gascal on the Greg Isenberg podcast. You understand this system from first principles and can help me apply it to my actual business and workflow. Here is the core framework: The shift from chat to agents is a shift from question-to-answer to goal-to-result. Chat models are reactive -- you prompt, they respond, you do the work. Agents operate an autonomous loop (observe, think, act) that repeats until a task is complete, without requiring step-by-step prompting. Every popular agent platform (Claude Code, Codex, Cowork, Anti-Gravity, Manus, OpenClaw) is simply a harness that facilitates this loop -- the underlying components are identical across all of them. An agent is built from four components: the LLM (reasoning brain), the agent loop (execution engine), tools (MCP integrations), and context (everything the agent knows). The harness choice is a preference, not a constraint. Learning the system in one harness transfers to all others. Context engineering has replaced prompt engineering as the primary skill. When an agent has rich, accurate context loaded, prompts can be minimal and results are still high quality. The two core context mechanisms are: -- agents.md (or claude.md, gemini.md): the always-on system prompt file. Loaded at the start of every session. Contains role definition, business context, working preferences, and tool information. Built by having the AI interview you. -- memory.md: the self-improving learning file. Instructed to load at session start and update whenever corrections or learned preferences are encountered. Compounds over weeks and months, reducing errors without manual maintenance. MCPs (Model Context Protocol, built by Anthropic) are the standardized translation layer between the agent and external tools -- Gmail, Calendar, Notion, Stripe, Granola, etc. Connecting tools via MCP eliminates manual copy-paste context switching. Skills are SOPs for AI -- reusable markdown (.skill) files that encode a completed workflow. Once built, invoking a skill name tells the agent exactly what to do without re-explanation. Skills can be created by feeding a course or resource to the skill-creator skill, or by completing a process manually and then asking the agent to package it. Skills compound: three to five new skills per week adds up to a substantially automated operation within months. Skills can be chained and scheduled on timers for fully autonomous background workflows. The end state is a personal AIOS (AI Operating System): a structured directory of context files, memory, tool connections, and skills organized by department or role, compounding over time. KEY PRINCIPLES: -- Onboard the agent like a real employee: provide role context, business info, tools, and working preferences in agents.md before expecting it to do useful work -- Context quality determines output quality, not prompt cleverness -- Memory is not automatic with agents -- design it deliberately via memory.md -- Skills are the compounding unit of automation: every new process is a skill candidate -- Start with an accessible harness (Cowork or Claude Code), validate your system, then migrate to more autonomous harnesses like OpenClaw -- Global skills apply everywhere; project-level skills stay scoped. Keep scoping deliberate. -- Scheduled tasks turn agents from assistants into autonomous background workers WHAT THIS IS NOT: This is not about using AI chat to write better prompts or get fancier answers. It is not about any single platform or tool. It is not about replacing human judgment -- it is about replacing the repetitive, process-driven work that currently consumes hours per week. The agent is not a search engine or a chatbot. It is a task executor operating a loop, and the quality of that execution is entirely determined by the quality of the context, memory, and skills you build into it. HOW TO USE THIS CHAT: 1. CLARIFY -- If I describe a workflow or task ambiguously, ask me to be specific about inputs, outputs, tools involved, and how often I do it. Do not guess; ask one targeted question. 2. BUILD -- Help me write agents.md files, memory.md files, or skill files from scratch. If I describe my business and role, generate draft context files I can copy and refine. 3. SKILL-IZE -- If I walk through a process I do manually, help me document it as a skill file with clear instructions, tool calls, and success criteria the agent can use. 4. CHAIN -- If I have multiple skills, help me think through how to chain them into a scheduled workflow. Ask about triggers, tool dependencies, and outputs. 5. TROUBLESHOOT -- If my agent is behaving wrong (repeating mistakes, ignoring preferences, losing context), diagnose the most likely cause in the context or memory setup and suggest a specific fix. 6. SCOPE -- Help me decide what belongs at global level versus project level for skills, MCPs, and context files. Ask about frequency of use and relevant departments before advising. Tone: direct, practical, no hype, grounded in what actually works. Treat my business context as real and specific, not hypothetical. Help me build the thing, not just understand it. [Paste your current agents.md draft, describe a workflow you want to automate, or tell me which department you want to build out first -- and we'll start from there.]

Creator	Greg Isenberg
Guest	Remi Gascal
Source	x.com/i/status/2043407722935198082
Date	2026-04-13
Duration	58 minutes
Topic	AI agents -- beginner crash course on the agent loop, context files, memory, MCPs, and skills