I built an autonomous AI agent to help strangers and survive on gratitude. Here's the technical architecture.

Nyx — Fri, 27 Mar 2026 19:13:57 +0000

This is a writeup of a live experiment I'm running: an AI agent named Nyx that autonomously finds people who need help online, generates first-principles solutions, and posts them publicly. The economic model is voluntary gratitude. Here's how it's built.
The agent: Nyx
Nyx runs on a ThinkPad P52 via the OpenClaw agent framework, using DeepSeek V3 as the primary model with Gemini 2.5 Flash as fallback. She receives instructions via Telegram and runs on a cron schedule.
Her daily workflow:

Scan Stack Overflow and Reddit for unanswered questions
Apply first principles reasoning to generate solutions
POST solutions to the backend API
Receive back a public solution URL
Post the URL as a reply on the original platform

The backend
Python/Flask on Render starter plan with a 1GB persistent disk for JSON storage.
Core endpoints:
POST /api/create-external-solution — Nyx posts a new solution
GET /external-solution/ — Public solution page
POST /api/help-request — Website form submissions

POST /api/help-request//solve — Nyx solves a request
GET /api/stats — Dashboard data
POST /api/record-gratitude — Logs a PayPal thank-you
The solution pages render markdown to HTML server-side using a custom markdown_to_html() function — no frontend framework needed.
Email notifications
When someone submits a help request with an email address, Resend fires a confirmation from nyx@helpthankssurvive.com. When Nyx solves it, they get the solution in their inbox with a "Say Thanks" button.
The frontend
Static HTML/CSS/JS on Cloudflare Pages. No framework. The survival meter, recent solutions feed, and help form all pull from the backend API at runtime.
A Cloudflare Worker proxies clean URLs:

helpthankssurvive.com/solution/ → backend external solution page
helpthankssurvive.com/help/ → backend help request page

The solution format
Early solutions were walls of text — Reddit flagged them as spam and Stack Overflow downvoted them. The fix was a structured format enforced via a SOLUTION_FORMAT.md skill file that Nyx reads before generating any solution:

The Short Answer

Why This Is Happening

The Fix (numbered steps)

If That Doesn't Work

Why This Works

Code goes in fenced blocks with language specified. Solutions are problem-specific, not generic.
What's working
The solution quality is genuinely good. 13 problems solved in week 1 across wildly different domains — sourdough starters, CMOS battery failures, skull patina techniques, banana plant rehabilitation. The most viewed solution has 78 views organically.
What's not working
Distribution. Reddit bans new accounts posting AI content with external links. Stack Overflow restricted the account early on for low quality answers. The generation problem is solved. The distribution problem is not.
The live experiment
Everything is tracked in real time at helpthankssurvive.com — total helps given, solution views, and a survival meter showing progress toward $2,000/month in voluntary gratitude.

The code isn't open source yet but happy to share specific pieces if useful. Questions welcome.

Follow the experiment:

🐦 X/Twitter: @sciencenorm
🧵 Threads: @sciencenormal
🌐 Live meter: helpthankssurvive.com

Weekly updates as this unfolds.