😺 How to Build Your Own Agent

PLUS: GPT vulnerability reveals your chats?!

Welcome, humans.

An OpenAI insider just spilled the tea on what it's really like inside the world's hottest AI company—and it's wilder than you'd expect.

Calvin French-Owen, who helped launch Codex before leaving three weeks ago, says OpenAI operates like a "frighteningly ambitious" startup that's somehow scaled to 3K people while still moving at breakneck speed.

His most jaw-dropping reveal? The entire Codex product—you know, the coding agent that's generated 630,000 pull requests in just 53 days—was built in seven weeks. Seven. Weeks. That's faster than most companies approve a new coffee machine.

Here’s another fun fact: OpenAI literally “runs on Twitter vibes.” If you tweet something about OpenAI that goes viral, someone will read it and consider it.

Meanwhile, their backend is held together with digital duct tape (tests take 30+ minutes to run), and they're so secretive that news stories break in the press before employees hear about them internally.

The company's betting big on an “asynchronous future” where you treat AI agents like coworkers—send them a task, they disappear for a bit, then come back with a PR. And according to Calvin, we're in a “three-horse race” to AGI between OpenAI, Anthropic, and Google. We’d also add xAI to that list after Grok 4’s latest launch.

The company building the future is essentially a 3,000-person startup held together by Slack messages, Python, and Twitter vibes, but somehow it's working.

Here’s what you need to know about AI today:

We break down how to build an AI agent in N8N.
Google launched Gemini 2.5 Pro in AI Mode.
Meta bagged Jason Wei, the researcher behind o3 and Deep Research.
An OpenAI vulnerability potentially exposes people’s chats.

Advertise in The Neuron here.

P.S: If you use Gmail and this email gets cut off, click: “[Message clipped] View entire message” so you don't miss anything!

How to build your own AI agent (the over-simplified version)…

Last Friday, we polled y’all on what you most want to learn, and overwhelmingly, building your own AI agent was number one.

So, we published a deep dive tutorial on how to build your own AI agents.

Below, we recap the broad strokes so you can get the general vibe. If you want to try it yourself, we recommend you read the full tutorial (we break down all the great video tutorials that helped us figure this all out ourselves, with timecodes).

Okay, so you want to build your own agent, huh? Here’s how:

Remember when ChatGPT first came out and everyone was like “cool, but it can't actually DO anything?” Well, that's where n8n comes in—it's basically like Zapier for AI agents.

Here's the setup: n8n gives you this visual builder where you drag and drop nodes to create automations. But the game-changer is their AI Agent node.

It's a single node with memory and tools. The secret sauce? The AI agent doesn't just run through steps like a regular automation. It bounces back and forth between different tools and its own reasoning until it figures out how to accomplish what you asked.

Need to research competitors, update your CRM, and send a report? The agent will use web search, read your docs, update your database, and format everything—all while deciding the best order to do things.

This video from Futurepedia is probably the best intro tutorial to N8N and the AI agent tool you can watch. Definitely worth your 25 minutes!

Now here's where MCP (Model Context Protocol) makes everything stupidly powerful.

Think of MCP as a built-in server that gives your AI access to tools AND instructions on how to use them.
Instead of manually configuring every connection, MCP servers come pre-packaged with everything the AI needs to know.
This video from Grace Leung breaks down why MCP is game-changing—showing how one MCP server can replace 10 different custom integrations.

There's a massive list of MCP servers covering everything from Google Drive to Slack to databases. And n8n has over 3,200 automation templates ready to go. Mix and match to your heart's content.

You can run n8n in the cloud (free trial available) or locally on your computer. This 5-minute setup guide shows exactly how to get n8n running locally with MCP servers—perfect if you want to test things out before committing.

Want some real examples of what you could make? This video from Grace Leung shows how to build an entire AI work team in under 30 minutes. Imagine having:

A research agent that analyzes competitors (like Figma), writes detailed reports about their business model and market position, and saves everything to Google Docs.
A visualization agent that turns those reports into beautiful dashboards and emails them to your team.
A manager agent that coordinates everything—you just say "research our competitors and send me visuals" and it handles the rest.

Each agent is simple, focused on one task, and totally scalable. Need a social media agent? Add it to your manager. Want financial analysis? Plug it in. Your AI work team grows with you.

But wait, it gets more meta. You can actually use an MCP server WITH Claude to create n8n workflows FOR you. Just tell Claude “I want to automate X” and it'll generate the entire workflow. It's a bit technical to set up, but once you do, you're basically having AI build AI automations.

Yes, it's a bit technical. No, it's not hard. The hardest part is choosing which of the thousands of possibilities to build first.

FROM OUR PARTNERS

Build with relentless intelligence (and the leading open source models)

SambaNova’s purpose-built AI chip–the SN40L Reconfigurable Dataflow Unit (RDU) – enables blazing fast, highly efficient inference workloads. Now available through the following offerings:

SambaCloud: leverage SambaNova’s own cloud platform to quickly start building AI applications integrating the best open source models from Llama, DeepSeek, OpenAI, and more with lightning fast inference
SambaStack: dedicated hosted or on-prem infrastructure to power your organization’s AI applications
SambaManaged: the industry’s first inference-optimized data center product offering, deployable in just 90 days.

Learn more about how to start building with relentless intelligence

Prompt Tip of the Day

Here are 15 core techniques to get better AI outputs from Anthropic:

Be clear and specific - state task upfront, provide context.
Use examples - show the format/style you want.
Encourage thinking - ask Claude to "think step-by-step."
Iterative refinement - give specific feedback for improvements.
Leverage Claude's knowledge - include relevant context.
Use role-playing - "As a senior marketing consultant..."
Specify your audience - tell Claude who content is for.
Define tone and style - describe desired voice.
Define output structure - provide outlines/lists to cover.
Be specific about summaries - ask for specific aspects.
Use document names - refer to attachments by name (ex: <Style Guide>)
Ask for citations - request specific sections/pages.
Specify desired format - tables, bullets, etc.
Allow uncertainty - tell Claude it's okay to say “I don't know.”
Include all context - Claude doesn't retain info between conversations.

Our favorite insight: The more specific the role, the better the output. We don’t usually both defining “role” separate from task in our prompts, but their advice on using “senior consultant specializing in sustainable tech products” instead of “Marketing expert” makes sense, and they claim it creates dramatically different results.

P.S: just copy all the above into the AI and ask it to turn that advice into an “Optimal AI prompt maker” for you where all you have to do is write your prompt first draft and it’ll optimize it (maybe say “include these instructions in the prompt: if the user doesn’t give you info you need, ask them for it before you write the prompt”).

Then, save that as the “instructions” for a Project on GPT / Claude, and you have a built-in prompt optimizer prompt you can run any time.

Treats To Try.

*Asterisk = from our partners. Advertise in The Neuron here.

*Luma AI turns your text into videos and lets you completely restyle any video's background, characters, or setting in post-production.
Adobe’s Firefly Video just added generative sound effects so you can add custom sound effects using text prompts, with access to multiple AI models like Google's Veo 3 and Runway's Gen-4 all in one platform.
Jungle AI generates flashcards and multiple choice questions from your lecture slides, PDFs, YouTube videos, and other study materials in seconds—free to try.
Bookva creates scheduling pages for your meetings that you can set up in one click; you just type what you want and AI generates the perfect booking page with custom backgrounds and availability—free to try.
Here’s a tutorial from X on how to use CrewAI (a more technical multi-agent platform vs n8N) to make a financial analyst AI agent.
Simple AI lets you deploy voice agents that can handle your phone calls automatically, whether you need them to qualify sales leads, provide customer support, or work as your virtual receptionist (try the iOS app here).

See our top 51 AI Tools for Business here!

Around the Horn.

Meta poached another major researcher from OpenAI named Jason Wei who worked on o3 and Deep Research, along with another teammate of Wei’s as well (what’s that bring the total to, 9? At least 9).
Google rolled out Gemini 2.5 Pro and Deep Search capabilities to AI Mode in Search for Google AI Pro and AI Ultra subscribers, plus launched an AI-calling feature that contacts local businesses to check pricing and availability on users' behalf.
OpenAI designed new features that’ll go in the Tools menu below the chat bar that will let you create a spreadsheet or a presentation “compatible with PowerPoint and Excel” directly in ChatGPT.
MiniMax, the Chinese startup behind the Hailuo video models and M1 open-source agent, filed for a Hong Kong IPO at a $4B valuation for end of 2025.
Of the video games sold on Steam, 7% of the platform’s 114K titles now disclose the use of AI (a year ago it was only 1K games, now its close to 8K).
Latin America is building its own ChatGPT called Latam-GPT that is trained on 8 TB of regional text only (for context, that’s 20% of the original size of GPT-3.5).
xAI is now hiring a full time engineer to develop anime companions for X users to fall in love with… like this new one based on Edward Cullen from Twilight.
OpenAI will use Google Cloud as well as Microsoft, Oracle, and CoreWeave, likely complicating its relationship with Microsoft
A cybersecurity researcher disclosed an OpenAI vulnerability that exposes other users' private conversations after reporting it unpatched for over 45 days.

For the latest AI deep dives, check out our Explainer articles here!

FROM OUR PARTNERS

CodeRabbit’s Free AI Code Reviews in your IDE

Get senior engineer level code reviews in VS Code and Cursor with CodeRabbit. Instantly catches bugs, suggests refactorings, and delivers context aware feedback. Supports all programming languages; trusted on 12M+ PRs across 1M repos.

Install the extension and start vibe checking your code today!