Building AI Agents that actually work (Full Course)

Main ideas & lessons

• AI is shifting from “chat” to “agents.”

  • Chat models are question → answer.
  • Agents are goal → result: they take a task, plan, execute steps, and iterate until completion.

• Agents operate using an “agent loop.”

  • The loop repeatedly does:
    • Observe (check context/workspace/available info)
    • Think (decide next steps, plans, decisions)
    • Act (use tools/write code/run tasks)
  • Completion depends on conditions/parameters specified in the prompt (e.g., “research 10 sources then create a PowerPoint”).

• An “agent” is composed of core elements

  • LLM (the brain/model) (examples mentioned: Claude Opus 4.6, GPT 5.4, Gemini 3)
  • Loop (keeps working; not one-shot)
  • Tool + context connections (so it can use your environment)
  • Agent harness/platform (the app that facilitates the loop across tools and context; examples mentioned: Claude Code, Codex, Antigravity, Co-work, Open Claw, Manas)

• The big practical leap is context + memory setup

  • In chat, models often have automatic memory you can’t fully control.
  • In agents, you typically must provide explicit context files (e.g., agents.md).
  • Optionally, you can add a manual long-term memory system (e.g., memory.md) so preferences persist across sessions.

• “Context engineering” replaces “prompt engineering.”

  • Instead of complicated prompts, you load the agent with rich business context so tasks can be simple (e.g., “write me a cold email”).

• Tools are connected via MCP (Model Context Protocol)

  • MCP acts as a translator/standard connector between the agent (speaks English) and apps (use their own “languages,” like Gmail/Notion/etc.).
  • Most agent harnesses support connecting tools through connectors/skills/settings using MCP.

• Skills are SOPs for AI (standard operating procedures)

  • A skill packages a repeated workflow into a reusable “dot-skill” markdown process (e.g., skill.md / .dot skill).
  • Purpose:
    • reduce repeated work
    • keep formatting consistent
    • avoid cluttering persistent memory with tactical details

• Skills + schedules create automation workflows

  • Harnesses introduce scheduled tasks (e.g., run a “daily brief” skill every morning).
  • This compounds time savings daily/weekly.

• Proposed “AI OS” future workflow

  • The speaker envisions a central workspace (an “AI OS”) where agents manage departments using connected tools and stored markdown context.
  • Emphasis: stop switching between apps; instead operate in a central agent environment.

• Security should be handled by scoping permissions

  • Treat tool access like granting privileges to an employee.
  • Use read-only where possible; limit what an agent can do if compromised.

---

Methodology / step-by-step instructions (as taught)

Build an AI agent workflow (from beginner to working department agent)

1. Choose a goal the agent will achieve

   • Example: “Build a minimalist portfolio site for Greg Eisenberg.”
   • Provide clear completion criteria in the prompt (e.g., how many sources, output format).

2. Understand the agent loop (observe → think → act)

   • Expect the agent to:
     • check workspace/files it can use
     • do research or planning when needed
     • generate outputs (code, drafts)
     • iterate until the “task complete” condition is met

3. Provide onboard context via an agents.md-type file

   • Purpose: ensure the agent knows who you are and how your business works.
   • Include:
     • role definition (e.g., “you are my executive assistant”)
     • business overview
     • tools used and how to use them
     • working preferences
     • ideal customer profile / target audience
     • tone and formatting preferences
   • File behavior described:
     • New sessions load this context automatically (when named/configured correctly).

4. (Optional but recommended) Add persistent preference memory via memory.md

   • Problem addressed: agents may forget updates across sessions unless you explicitly store them.
   • Approach taught:
     • Add instructions in agents.md telling the agent to:
       • read memory.md
       • update it when new preferences are learned
       • replace outdated info in-place

5. Connect tools using MCP

   • High-level steps:
     • open the harness (e.g., Co-work / Claude Code / Codex / Manas)
     • browse connectors
     • sign into apps (Gmail, Calendar, Notion, Granola, Stripe, etc.)
     • connect via MCP so the agent can call those tools

6. Create reusable skills for repeated workflows (SOPs)

   • Define what counts as a repeated process you’ll do again (e.g., proposals, cold email formats, referral emails, daily brief, meeting prep).
   • Two creation methods mentioned:
     • Skill creator skill using a course/transcript
       • Upload/transcribe a material (e.g., “viral hooks course”) and instruct the harness to convert it into a skill.
     • Skill creator skill after doing the process once
       • Complete the workflow manually once, then ask it to package that process as a skill.

7. Schedule skills for automation

   • Example: run “daily brief skill” every morning at a specified time.

8. Chain skills into multi-step workflows

   • Example concept:
     • morning brief skill checks for meetings/podcasts
     • trigger podcast research skill
     • prepare talking points
     • then send an email or create deliverables

---

Build and use an “Executive Assistant” example (as demonstrated)

• Create a local folder/project (e.g., executive assistant).
• Start with a simple task: “Write me a cold email and send it.”
• If the folder has no context:
  • the agent will ask clarifying questions because it doesn’t know your business.
• Add the proper context file (rename to match harness expectation, e.g., agents.md / Claude.md / Gemini.md).
• After adding context:
  • the agent can draft cold emails with correct tone and targeting.
• Add memory.md if you want cross-session learning (e.g., signature preferences).

---

Examples referenced in the video

• Agent loop demo

  • Build a minimalist portfolio site for a named person.
  • Agents:
    • research the person
    • write code
    • run/preview locally
    • screenshot and verify completion

• Cold email workflow

  • Agent drafts cold email based on onboarding context.
  • Suggested outreach use case: bulk pre-made websites offered as an off-the-shelf product.

• Tool-driven workflow

  • Summarize inbox from today using connected email + notes + payments + project management tools.
  • Example flow:
    • summarize inbox / review meeting notes
    • draft email proposal
    • create a Stripe payment link
    • create a Notion project
    • send via Gmail integration

• Skills examples

  • “Viral hooks” skill created from a course transcript.
  • “Ads analyst” skill
    • scrape competitor ads library
    • screenshot landing pages
    • analyze copy + creatives
    • output master report
  • “Sebastian referral skill”
    • when asked to refer someone, it knows email and steps to connect an introduced prospect

---

Security guidance (what was said)

• Think of security as scoping what access the agent has.
• By default, major harnesses are described as secure due to company resources.
• Practical risk control:
  • limit tool permissions
  • prefer read-only access for sensitive systems
  • minimize worst-case damage if something is compromised
• Agent permissions are likened to giving a risky employee privileges (e.g., managing ad budgets).

---

Speakers / sources featured (and identified)

Speakers

• Greg (host / interviewer; Greg Eisenberg referenced)
• Remy Gaskill (guest; course instructor)
• Cody Schneider (mentioned as someone whose tweet/prediction was referenced)

Other sources / entities mentioned (not speaking)

• MCP (Model Context Protocol) — attributed to Anthropic in the explanation
• Agent harness / platforms mentioned: Claude Code, Codex, Antigravity, Co-work, Open Claw, Manas
• Model examples mentioned: Claude Opus 4.6, GPT 5.4, Gemini 3
• Tools connected via MCP (examples): Gmail, Google Calendar, Granola (meeting notes), Notion, Stripe
• Research tool mentioned: Perplexity
• Transcription tools mentioned: Monologue, Whisper (and “Whisper flows”)
• Guest video mention: “Ross Mac” podcast/video (referenced for MCP explanation)