Agent Evan

---

the obsession

Michel and I want to be at the edge of what's possible with AI. We stay up until 3 AM reading new papers the night they drop. When DeepSeek dropped DeepSeek-OCR, we spent an entire weekend tearing it apart layer by layer. When Claude gets an update, we're stress-testing it within hours.

So when we walked into HackRice, we were thinking about agents. Real agents. The kind that actually do things.

The initial spark for Agent Evan

---

the problem

Every AI tool right now is a conversation. You ask it something, it responds with text. Maybe it writes you some code. Maybe it gives you instructions. But at the end of the day, you're still the one who has to do everything.

Copy the code into your terminal. Navigate to the right folder. Run the commands. Export the file. Send the email. You're the hands. The AI is just the mouth.

We wanted to flip that completely.

---

the solution

Evan just does things.

Evan checking calendars and writing emails autonomously

Tell Evan to create a presentation about your research with proper formatting and export it as a PDF. Done. No touching PowerPoint.

Tell Evan to find every PDF in your Downloads folder from the last month, merge them, and save it to your Desktop. Done. You watch it happen.

Tell Evan to scrape that website, clean the data, throw it into a spreadsheet, and generate a chart. Done. The file appears.

Tell Evan to start a training job, monitor the logs, and let you know when accuracy hits 90%. Done. You go do something else.

Evan running Jupyter Notebooks remotely

This works because Evan has access to everything on your computer. Your file system. Your terminal. Your applications. Your browser. It can run code, process images, generate documents, control apps, monitor processes. Anything you can do sitting at your desk, Evan can do too.

And you don't need to be at your desk. Evan takes commands from your phone. You could be on the bus, at lunch, in another country. You type what you want, your computer executes.

The difference between Evan and every other AI tool is simple. ChatGPT tells you how to do things. Evan does them.

Agent Evan using tools and creating documents

---

the 36 hours

How Agent Evan works

We had to build three systems that work together perfectly: a mobile app, a server, and a desktop agent. In 36 hours. With four people.

The mobile app runs native iOS in Swift and SwiftUI. We used Apple's new Liquid Glass design so it feels like it belongs on your phone. Cartesia's Sonic model handles speech-to-text in near real-time. You talk, Evan understands, all in milliseconds.

The server runs on Cloudflare Workers with Supabase S3 for storage. WebSockets keep your phone and desktop connected instantly. When Evan creates something big like a presentation or a processed video, it syncs through the cloud so you can access it anywhere.

The desktop agent is where the real magic happens. Built in Python with Claude as the brain. Every agent runs inside its own sandboxed Docker container for security and isolation. This also lets us run multiple tasks at the same time without anything breaking.

Our debugging interface for tools

We built over 130 tools. Web search and fetch. Shell commands. File operations across your entire system. Mail and calendar access. Browser control through Playwright. Jupyter notebook execution. LaTeX and Pandoc for document generation. LibreOffice automation. OpenCV for image processing. PyTesseract for OCR. Screen capture. Clipboard access. Process monitoring. Application launching. And dozens more.

130 tools. All available to the model at once. All the time.

That last part is important. Most agent frameworks try to be clever by hiding and showing tools based on context. We did the opposite. We give Claude every single tool and let it figure out what to use. The model knows best.

During development, when certain tools broke, Evan would notice and find alternative approaches on its own. We never programmed that behavior. The model figured it out because we gave it the freedom to think.

---

the hard parts

Security versus access was brutal. We needed the agent to control your actual computer while keeping everything sandboxed and safe. Getting a Linux VM inside Docker to expose exactly the right interfaces took hours of debugging.

Real-time communication is fragile. WebSocket connections drop. Networks hiccup. Your phone and desktop need to stay synced through all of it. We spent a huge chunk of time making the connection layer robust.

Voice latency matters. When you speak a command, you expect near-instant response. Every millisecond in the pipeline from speech to execution to confirmation had to be optimized.

Prompt engineering for tool selection was its own beast. Getting Claude to reliably choose the right combination from 130+ tools required endless testing and iteration.

---

the win

We won the HackRice 2025 AI Track.

We submitted at 8 AM. Completely exhausted. We passed out.

Funny story: we were so sleep deprived we didn't even show up to the final award ceremony. We only found out we won because a friend texted us.

THANK YOU, ERIC!!!

---

what we shipped

An AI agent with full desktop access. 130+ tools for the model to choose from. Intelligent tool selection and chaining. Concurrent execution for complex multi-step workflows. Real-time monitoring and updates. Document and media creation. Voice and text commands from mobile. Sub-second response times.

You tell it what you want. It does it.

---

what's next

Evan today is a small step toward something bigger.

Short term we want to expand beyond macOS, support more models, and polish the rough edges we didn't have time to fix in 36 hours.

Long term, as reasoning models get more capable, Evan will handle increasingly complex workloads. The goal is simple. You focus on the ideas. Evan brings them to life.

---

Agent Evan

Built at HackRice 2025 by Michel Guo, Hemesh Chadalavada, Ethan Harjabrata, and Demetris Chrysostomou