Letta Code Action - YouTube Video Script#

Target Length: 6-8 minutes
Style: Talking points (not word-for-word)

HOOK (0:00 - 0:30)#

Key points:

• AI coding assistants today are stateless - they forget everything
• Imagine an AI that actually remembers your past conversations
• That's what Letta Code Action does - persistent AI agents in GitHub

Suggested opener:

"What if your GitHub issues had an AI that remembered everything you ever talked about?"

THE PROBLEM (0:30 - 1:30)#

Points to hit:

• Every time you use Copilot in issues, it starts fresh
• You waste time re-explaining context constantly
• "Remember when we discussed the auth system?" "No, what auth system?"
• Traditional bots are stateless by design

Visual: Show a generic AI response that clearly doesn't know context

INTRODUCING LETTA CODE ACTION (1:30 - 2:30)#

Key points:

• Open source GitHub Action
• Mention @letta-code in any issue or PR
• THE MAGIC: Same agent resumes across conversations
• Agent ID stored in comment metadata - it knows who it is

Visual:

• Show the repo briefly (github.com/letta-ai/letta-code-action)
• Point out the agent link in a comment footer

DEMO PART 1: First Contact (2:30 - 3:30)#

What to do:

1. Show this demo repo (demo-todo-api)
2. Quick tour: "Simple Todo API, Express, TypeScript"
3. Create a new issue
4. Type: @letta-code can you explain how the authentication middleware works in this API?
5. Wait for response (or use pre-recorded)

Points to make:

• Agent explores the codebase autonomously
• Gives a real answer based on the actual code
• Notice the agent link in the footer

DEMO PART 2: The Memory Test (3:30 - 4:30)#

What to do:

1. In the SAME issue, add a new comment
2. Type: @letta-code I think there's a bug in there. Can you find it?
3. Show the response

Points to make:

• Notice: "I didn't say WHAT or WHERE"
• Agent remembers we were talking about auth middleware
• It finds the case-sensitivity bug
• This is the persistence in action

DEMO PART 3: Building on Context (4:30 - 5:30)#

What to do:

1. Still in the same issue
2. Type: @letta-code now can you add input validation to the POST endpoint?
3. Show the agent's response/PR creation

Points to make:

• Agent knows the full context now
• Doesn't need re-explanation of the codebase
• Can create PRs directly with fixes
• Memory compounds over time

HOW IT WORKS (5:30 - 6:15)#

Quick technical overview:

• Agent ID embedded in HTML comment (show example)
• When you mention again, it extracts the ID
• Resumes the same agent with full memory
• Built on Letta's stateful agent framework

Keep it brief - don't go too deep, link to docs

SETUP (6:15 - 6:45)#

Show the workflow file:

- uses: letta-ai/letta-code-action@main
  with:
    letta_api_key: ${{ secrets.LETTA_API_KEY }}

Points:

• That's basically it
• Get API key from Letta
• Add to repo secrets
• Works on issues and PRs

CLOSING (6:45 - 7:15)#

Key takeaways:

• Persistent AI agents that remember context
• Open source - see exactly how it works
• This is what agentic AI actually looks like

CTA:

• Link in description
• Star the repo
• Try it on your own projects
• Subscribe for more AI coding tool breakdowns

B-ROLL SUGGESTIONS#

• GitHub issues interface
• Code scrolling (auth.ts, todos.ts)
• Terminal running the API
• Letta logo/branding
• Agent link in comment footer

PREPARED COMMENTS FOR DEMO#

Copy-paste these during recording:

Comment 1:

@letta-code can you explain how the authentication middleware works in this API?

Comment 2:

@letta-code I think there might be a bug in the auth middleware. Can you review it and find any issues?

Comment 3:

@letta-code nice find! Can you also check if there are any input validation issues in the routes?

Comment 4 (optional - PR creation):

@letta-code can you create a PR that fixes the auth bug and adds input validation?

KNOWN "BUGS" FOR DEMO#

The agent should find these:

1. Auth middleware (src/middleware/auth.ts):

   • Case-sensitive token comparison
   • Should use .toLowerCase() for comparison

2. Todo routes (src/routes/todos.ts):

   • No input validation on POST /todos
   • Title could be empty, undefined, or wrong type
   • No try/catch error handling

TIMING CHECKLIST#

• Hook grabs attention in first 10 seconds
• Problem explanation under 1 minute
• Demo is the meat (3+ minutes)
• Setup section is quick (<30 seconds)
• Total runtime: 6-8 minutes