Six days ago, at 4:45 in the morning, I typed "What exactly is an AI agent?" into a Claude chat window. I wasn't being rhetorical. I genuinely didn't know.
I'm a video editor. 13 years in feature animation, Blue Sky Studios, Netflix Animation. Ferdinand, The Peanuts Movie, Rio 2, Nimona. My entire career has been managing complex post-production pipelines, coordinating across departments, and shipping projects under impossible deadlines. I've never written production code. I don't have a CS degree. Python was something I'd seen in memes.
Today I'm open-sourcing First Contact, a personal AI agent with 450+ tests, 4 interfaces, multi-model routing, a user model that learns from your conversations, and a job search pipeline. I built it in 6 days. And I didn't write a single line of code.
What I Built
Here's what First Contact actually is: a personal AI agent that runs on your machine, manages your workflow, and gets smarter about you over time.
The numbers: 467 tests, 46 modules, 28 tools, 4 interfaces (terminal, web GUI, Discord bot, Telegram bot), 3 LLM providers, full document support including PDF/DOCX/XLSX read and creation, image input across every interface, and a persistent user model that learns from your conversations without you doing anything.
Is it perfect? No. Are there bugs? Absolutely — I'm squashing them right now as I write this. But that's kind of the point. A year ago, a non-developer building something like this was impossible. Six months ago it was impractical. Today I'm debugging my own AI agent at midnight and having the time of my life doing it.
The bar to entry for building real software has dropped through the floor. If you have an idea and you're willing to sit with it, you can build it. Not a toy demo. A real thing with real tests that real people can install and use.
How I Actually Built It
I want to be completely transparent about this: I didn't write the code. Claude Code wrote the code. Every single line. That's called vibe coding and I'm not going to pretend otherwise.
But here's what I did do. I architected every decision. Which models handle which tasks — Haiku for cheap extraction work, Sonnet for reasoning, Opus for high-quality output like cover letters. Where to cache prompts to cut costs. How to split the user profile into two tiers so you're not burning tokens on irrelevant context every conversation. Why the security model should be structural, not just prompt-level. What to build, what to skip, and when to ship instead of polish.
I direct the AI the same way I used to direct an edit. I don't hand-animate every frame. I tell the team what the final product should look like, I review the output, I catch what's wrong, and I push until it's right. The tool changed. The skill didn't.
Do I understand every line of Python? No. But I read the code. I test the features. When something breaks, I describe the problem and work through the fix. And I'm learning more every day, not because I have to, but because I genuinely want to. This is the most fun I've had working in years.
Architecture Decisions I'm Proud Of
Security-first. This wasn't an afterthought. The security model is structural — capabilities are defined by the tool registry, not by prompts the model can talk its way around. Email is draft-only, the agent can never send on your behalf. File access is sandboxed. External data is never trusted as instructions. When I saw competitors getting flagged for data exfiltration through third-party plugins, I knew security had to be built into the architecture, not bolted on.
The agent has a heartbeat. First Contact doesn't just respond when you talk to it. A background daemon runs scheduled tasks: daily briefings, email monitoring, job scanning every 12 hours, reminders, and a proactive insights system that cross-references your calendar, email, tasks, and job applications to surface connections you'd miss. It told me a company I applied to posted a better-matching role before I'd even seen it.
It learns about you. After every conversation, the agent extracts facts, preferences, and patterns about you automatically. No setup, no manual profile editing. Over time it builds a model of who you are and what you care about. High-confidence stuff like your name and preferences are always in context. Everything else is filtered by relevance to what you're actually talking about so it's not wasting tokens. Nobody else does automatic extraction from conversations.
Job search pipeline. This is the feature that made me build the whole thing. Automated scanning across job boards, AI-powered fit assessment, cover letter generation, application tracking. I looked at every major open source agent out there. None of them have this. It's the thing that makes First Contact a daily-use tool, not a tech demo.
Cost optimization. Most AI tools drain your API budget without telling you. First Contact routes cheap tasks to cheap models, caches prompts across three levels, batches job assessments at 50% off, and tracks your cost per-turn so you always know what you're spending. I'm a freelancer, I built this because I needed it to be affordable to actually use every day.
What I Learned
A week ago I didn't know what an AI agent was. Now I have one running on my machine that checks my email, scans for jobs, and learns what I care about while I sleep. If that doesn't tell you the bar has moved, I don't know what does.
The biggest lesson wasn't technical. It's that the skill that matters now isn't coding, it's systems thinking. Knowing what to build, how the pieces connect, where to invest quality and where to cut costs. I spent 13 years developing that skill in an edit bay without realizing it was transferable. It is.
The other thing I learned: there's no single best AI model anymore. There's a best model for each problem. Cheap models for simple extraction. Mid-tier models for reasoning. Top-tier models for the work that actually matters. The people who figure out that routing, whether manually or by building tools that do it for them, are going to have a serious edge. I built the tool. Now I'm sharing it.
Check It Out
First Contact is MIT open source. Install it, break it, tell me what's wrong.
GitHub: github.com/slomei/first-contact
I'm also looking for work — AI trainer roles, video editing, anything where systems thinking and a bias toward shipping matter. My site is slomei.com. Say hi.
Top comments (0)