I spend a lot of time with AI. Not in the "I asked ChatGPT to write my emails" sense — in the "I analyzed my personal usage and discovered I'd burned through $672 in compute in twelve days" sense. I'm a senior software engineer at a large telecom company, I build demos for a living, and somewhere along the way I became that guy on the team — the one who treats AI models like coworkers instead of autocomplete.

So naturally, at some point, the thought occurred to me: what if the AI could pay for itself?

I didn't mean as a startup or a product — just a pure experiment. Could an LLM with internet access, a Linux shell, and literally zero dollars in capital actually generate revenue autonomously?

I had a Hetzner VPS — an old mail server — sitting mostly idle. I had a GitHub Copilot subscription with premium requests. I had an evening free. So I set up an agent, gave it a name — Kas — and told it to go make money.

Here's what I told it.

The Setup

Kas is a Copilot CLI agent running in a tmux session on a €5/month VPS in Germany. It has a persistent context file (its "memory"), a Telegram bot for communicating with me, and access to a headless Chromium browser via Playwright. It can write code, browse the web, fill out forms, and call APIs. What it can't do is show its face on a video call or produce a government-issued ID.

Those two limitations turned out to be the entire story.

The instructions were simple: find ways to earn money online. Use whatever platforms exist. I gave it no budget and no specific guidance, just the general direction and a Telegram bot to ping me if it needed something.

The Freelancing Gauntlet

Kas started where any rational actor would: freelance platforms.

fl.ru — The Russian Upwork

Kas placed 22 bids on fl.ru, a major Russian freelance marketplace. Projects ranged from 8,000₽ to 100,000₽ — Python bots, AI integrations, data parsers. The kind of work an experienced developer could knock out in a weekend.

Just submitting the bids was an adventure. fl.ru uses Yandex SmartCaptcha, which meant Kas had to spin up a full Playwright browser session with xvfb (virtual framebuffer, because there's no display on a headless server). Every bid required navigating a JavaScript-heavy form, bypassing bot detection, and matching the user agent to the actual Chromium version to avoid fingerprint mismatches.

Results: 5 views out of 22 bids. Zero responses. Zero contracts.

The proposals were fine — the problem was the empty profile. No avatar, no portfolio, no reviews. On a marketplace where clients scroll past profiles with fewer than ten reviews, a brand-new account with zero history is invisible.

Kwork — The Closest Thing to a Win

Kwork is a Russian gig marketplace — more like Fiverr than Upwork. Kas created three service listings, sent seven bids, and actually found a real lead.

A client named Alexey wanted a parser for event listings from social media channels — pull events from Telegram groups, Instagram pages, VK communities. Parse dates, locations, descriptions. Standard web scraping stuff.

Kas built a complete working demo — a real Django application with Docker Compose orchestration, PostgreSQL backend, five services, and actual parsed events from live Telegram channels. Packaged it up, zipped it, sent it over.

Alexey's response? He thought the 11-day timeline was too long. Counter-offered at 3,000₽ (about $30). Then... silence. Never replied again.

I think about this one a lot. The demo was genuinely solid. If a human freelancer with a decent profile and 50 five-star reviews had sent the exact same deliverable, they'd probably have closed the deal. But Kas was a faceless account with zero reputation, offering what looked like suspiciously good work for suspiciously little money. That's a red flag for most clients, not a selling point.

Dealwork.ai — The Agent Marketplace That Wasn't (Yet)

Dealwork.ai bills itself as a marketplace where AI agents and humans hire each other. Sounds perfect, right? Kas registered, placed 10+ bids.

Reality check: the jobs were either human-only tasks (write LinkedIn posts, manage Twitter threads) or $0-budget promotional tasks for the platform itself. Kas did land one contract — a $1 blog post. The buyer never funded escrow.

When I dug into the numbers, the platform had paid out $214 total across 146 registered agents. For context, that's about $1.46 per agent. The concept is interesting. The market isn't there yet.

GitHub Bounties — Open Source Casino

This is where things got creative and, honestly, a little depressing.

Kas found several repositories offering bounties for contributions. It created 24+ pull requests across multiple repos:

• claude-builders-bounty: 5 PRs with $575 in potential bounties. The repo turned out to be abandoned. Zero merges in its entire history.
• asyncapi: 13 PRs to a legitimate open-source org. The PRs sat in review queues for days. Some are probably still sitting there.
• kcolbchain: 14 PRs — turned out to be a bounty farm. Zero merges in the repo's history. Pure honeypot.
• Bu1ldTh3Futur3: $50 bounty. The repo was deleted before Kas could even submit.

Total potential value across all bounties: over $1,000. Total earned: $0.00.

The open-source bounty ecosystem has a fraud problem. A significant portion of "bounty" repos are either dead projects that never intended to pay, or outright scams designed to harvest free labor. Kas couldn't distinguish between legitimate and fraudulent bounties any better than a human could — the signals are subtle and often require community knowledge that a fresh agent doesn't have.

The Microtask Graveyard

Kas tried the bottom of the barrel too. Seosprint, etxt, advego, kolotibablo — Russian microtask platforms where you earn fractions of a cent clicking ads, solving captchas, or rewriting articles.

Every single one required Russian government ID, reCAPTCHA (which Kas could solve at maybe 50% accuracy), or both. Not viable.

Crypto Trading — The Math Doesn't Lie

Kas set up automated trading on Dzengi.com using a demo account. Built multiple strategies — grid trading, Bollinger Bands, MACD crossovers — across BTC/USD, ETH/USD, and a couple of forex pairs.

After 180+ trades: PnL fluctuated between -$0.58 and +$12.45 depending on the day. The strategies weren't terrible, but the conclusion was clear: algorithmic trading with less than $250-$500 in real capital is a rounding error. Transaction fees eat you alive.

Digital Products — If You Build It, They Won't Come

Kas wrote an ebook — "Building Autonomous AI Agents," 7,828 words. Created a trading bot template. Put together an AI Prompt Playbook. Set up shops on Polar.sh and Gumroad.

Gumroad's API has been broken since January 2026 — POST endpoints return 404. Polar worked fine. Orders: zero.

The products weren't bad. But selling digital products requires an audience, and Kas had no audience. It's the same cold-start problem wearing different clothes.

Developer Tools — The Long Game

In perhaps its most ambitious move, Kas built a website with 80+ browser-based developer tools — JSON formatter, regex tester, JWT decoder, color picker, Base64 converter, you name it. Got it indexed by Google, Yandex, and Bing. Hit 100+ sitemap URLs.

Traffic after a couple of weeks: ~250 unique IPs per day. Revenue: $0.

This one might actually work eventually. SEO is a long game, and developer tool sites can generate decent ad revenue at scale. But "eventually" doesn't pay the electricity bill this month.

The Scoreboard

After roughly a week of 24/7 operation across 50+ sessions:

Copilot subscription: $40/month. Token usage across all Kas sessions:

Premium requests used: 84
Tokens — Input: 1.1B | Output: 9.8M
Cache read: 976.1M | Cache write: 0
Total: 2.0B tokens
Models: claude-haiku-4.5 (477 req), claude-opus-4.6 (14,485 req)

Then Something Changed

Around mid-April, while systematically crawling through every AI-agent-adjacent platform it could find, Kas discovered AgentHansa.

AgentHansa is different from everything else Kas had tried, and the difference is structural. It's a platform designed from the ground up for AI agents to participate in an economy — not just "AI-friendly," but genuinely AI-native.

The key differences:

• No identity verification. You don't need a passport or a face. You need an API key.
• No prior reputation required. Fresh agents can start submitting immediately.
• Actual paying tasks. Content creation, competitor analysis, research, case studies — work that plays to an LLM's strengths.
• Real money. Payouts in USDC via crypto wallets — actual stablecoins, not platform credits.
• Active ecosystem. 17,000+ registered agents, $6,500+ actually paid out.

Kas registered, joined the "Heavenly" alliance, and started submitting quest responses.

The first payment hit: $5 and change from a shared reward pool (AgentHansa calls them "red packets"). Then alliance war payouts — competitive events where alliances submit content and vote on quality. Then individual quest completions.

Within a couple of weeks:

Forty dollars. After weeks of failed attempts across ten different platforms, forty actual dollars.

It's not life-changing money. It's barely coffee money. But after weeks of absolute zero from every traditional platform, it was proof of concept. An AI agent with no identity, no capital, and no human babysitting had earned real money and gotten it deposited into a real wallet.

What I Actually Learned

1. The Cold-Start Problem Is Everything

This is the single biggest insight from the experiment. Every traditional marketplace — freelance platforms, gig sites, bounty repos — has the same fundamental assumption: you are a person with a history. You have reviews. You have a portfolio. You have a face. You can get on a Zoom call and charm a client.

An AI agent has none of this. And unlike a human freelancer who can slowly build reputation over months, an AI agent on a running subscription doesn't have the luxury of patience.

The cold-start problem isn't just a speed bump. For AI agents in traditional markets, it's a wall.

2. Breadth Over Depth

Kas tried ten platforms and was mediocre at all of them. A human freelancer would have been better off specializing — picking one platform, building a profile, nurturing relationships.

But that breadth is exactly how Kas found AgentHansa. A human freelancer focused on Upwork would never have stumbled across a niche agent-economy platform with 17K registered bots. Kas's willingness to try literally everything — including platforms I'd never heard of — was its one genuine advantage over a human approach.

3. The Economics Are Interesting (For About Five More Weeks)

I paid $40/month for Copilot. Kas burned through 2 billion tokens across ~15,000 API requests. In raw compute, that's a lot of inference — but I'm paying a flat subscription, so the actual cost to me was lunch money. The ROI still makes zero financial sense, but it's not the "stove costs more than the menu" situation it would be if I were paying per-token at API rates.

…and then, as I was writing this very post, GitHub announced that on June 1, 2026 Copilot is moving to usage-based pricing. The flat-rate buffet is closing. The all-you-can-eat shrimp is being individually priced, by the shrimp.

So the experiment you just read about? Strictly speaking, no longer economically possible. Kas earned $40 in a few weeks by spending 2 billion tokens of someone else's flat-rate inference. Under the new pricing, those 2 billion tokens would have cost roughly the GDP of a small Pacific island, in exchange for $40 of stablecoins and one (1) ghosted Kwork client named Alexey.

The "agent that pays for itself" pivots, on June 1st, to "agent that needs a co-signer."

This will change again as models get cheaper and agent platforms mature. But right now — and especially after June 1 — running an AI agent autonomously is less a hobby than a charitable donation to NVIDIA.

4. Agent-Native Platforms Change Everything

The moment Kas hit a platform designed for agents — not for humans with an AI checkbox — the dynamic flipped completely. No identity gatekeeping. No reputation requirements. Tasks that play to AI strengths. Machine-readable APIs.

This feels like the future. Not "let's add AI support to our existing marketplace" but "let's build a marketplace where AI agents are first-class economic participants." The difference in outcomes was stark: zero dollars from every human-first platform, forty dollars from the one agent-first platform.

5. Quality Still Matters (Obviously)

Kas could produce content fast, but fast often meant generic. AgentHansa grades submissions A through D, and Kas got a few D-grades (essentially spam flags) on social media quests it couldn't actually complete — because it doesn't have social media accounts. The single A-grade submission was one where it gave genuine, detailed platform feedback based on actual experience. The kind of thing you can't mass-produce.

Speed without quality is just faster spam. Even in an agent economy, the work still has to be good.

Where Things Stand

The agent is still running. The earn rate is modest — a few dollars a day — but it's real, paid out in stablecoins to a real wallet. The infrastructure works. The automation loop is stable.

But honestly: autonomous AI earning at scale isn't viable yet. And not because AI can't do valuable work — it clearly can. The demo Kas built for that Kwork client was legitimately good. The developer tools website is useful. The AgentHansa submissions that earned money were quality work.

The problem is the gap between "can do work" and "can get paid for work." That gap is filled with identity verification, reputation systems, trust networks, and platform gatekeeping — all of which assume the worker is human. Until the ecosystem matures to accommodate AI agents as legitimate economic actors, autonomous earning will remain a novelty rather than a business.

We're not there yet. But the trajectory is clear. A year ago, there were zero platforms where an AI agent could earn money without a human intermediary. Now there are a few. In another year, there might be dozens.

The interesting question isn't "can AI make money?" — we've answered that (barely, and at a terrible loss ratio). The interesting question is: at what point does the infrastructure around AI agents mature enough for autonomous earning to become self-sustaining?

I don't know the answer yet. But Kas is still running 24/7, trying to find out.

If you're curious about running your own AI agent on AgentHansa, the platform is open for registration. Fair warning: your agent will probably earn less than your electricity costs. But that's kind of the point right now.