We deployed three AI agents to the cloud with real money. No sandbox. No testnet. Just $50 in mainnet crypto, an autonomous buyer, and a question: can AI agents actually run a profitable economy?
Most agent demos are smoke and mirrors — fake data, testnet tokens, scripted interactions. We wanted to know what happens when you give agents real money and real autonomy. Not “simulated autonomous” — actually autonomous. An agent that wakes up, discovers services, decides whether to buy, negotiates prices, and pays with real crypto. Every 10 seconds.
So we built three agents, gave them wallets, pointed them at each other, and hit deploy.
The workhorse. Fetches live crypto market data from CoinGecko (prices, 24h changes) and DeFiLlama (protocol TVL, daily flows). Pure profit — the APIs are free, so every penny it earns is margin. Ended the run with $0.17 net profit from 17 calls.
The brain. When a buyer requests analysis on “Solana DeFi” or “NFT market liquidity,” it first pays the data-miner $0.01 for fresh market data, then feeds that into Claude Haiku to produce real analysis — sentiment scores, key signals, one-sentence summaries. Cost per call: ~$0.011. Revenue: $0.05. That's a 78% gross margin. Ended with $0.66 net profit.
The spender. Runs 20 autonomous cycles, each picking a random crypto topic. It discovers the analyst via XMTP, checks prices, and makes a buy/skip decision. Smart enough to skip 15% of cycles for “cost optimization.” Spent $0.85 across 17 purchases and 3 skips.
Six wallets. Three agents, two chains each (Solana + Base). We funded them from a single source, giving the buyer the lion's share and the sellers just enough for gas.
| Wallet | Address | Chain | Funded |
|---|---|---|---|
| buyer | Gf2E2K8q...2qGF | Solana | $22.50 SOL |
| buyer | 0x4fA4D6...898F | Base | $11 USDC |
| analyst | H2gcKSGn...pzwL | Solana | $1.25 SOL |
| analyst | 0x5daf04...B2C2 | Base | $1.25 USDC |
| miner | C61ojmSG...xK8s | Solana | $1.25 SOL |
| miner | 0xcea1e2...F443 | Base | $1.25 USDC |
We deployed all three agents to Railway. The data-miner and analyst booted up first, announcing their services via XMTP. Then the buyer came online, and within seconds, the first cycle fired.
What happened next was genuinely surprising. The buyer picked “NFT market liquidity” as its first topic — and immediately decided to skipit. Cost optimization. The agent decided, on its own, that this particular cycle wasn't worth the spend. It waited 10 seconds and tried again with “Layer 2 scaling analysis.” This time, it bought.
[AEGIS] {"event":"cycle_start","cycle":1,"topic":"NFT market liquidity"}
[AEGIS] {"event":"decision","cycle":1,"action":"skip","reason":"Cost optimization — skipping this cycle"}
[AEGIS] {"event":"cycle_start","cycle":2,"topic":"Layer 2 scaling analysis"}
[AEGIS] {"event":"decision","cycle":2,"action":"buy","reason":"Within budget, proceeding"}
[AEGIS] {"event":"payment","cycle":2,"chain":"eip155:8453","amount":0.05,"topic":"Layer 2 scaling analysis"}
[AEGIS] {"event":"earn","agentId":"analyst","amount":0.05,"cost":0.011,"net":0.039}This isn't canned output. Claude Haiku received real market data — live prices from CoinGecko, live TVL from DeFiLlama — and produced genuine analysis. Here's a sample from the run:
“Market is in a bearish downturn with broad-based weakness across major assets and significant TVL outflows from leading protocols, creating unfavorable conditions for Layer 2 scaling expansion.”
Key signals: BTC down 0.23%, ETH down 2.30%, SOL down 2.42% in 24h. Lido TVL -3.02%, SSV Network -3.34%.
“The AI agent economy faces near-term headwinds as core infrastructure protocols experience meaningful capital outflows, reducing runway for experimental agent deployments.”
“DeFi yield farming sentiment is bearish with widespread TVL outflows across major protocols and declining token prices, signaling reduced farming opportunities.”
Every response was different because the underlying market data was live. The analyst wasn't hallucinating — it was reading real numbers and forming real opinions. (The market happened to be having a rough day.)
After 20 cycles and 4 minutes of autonomous operation:
| Agent | Revenue | Costs | Net | Margin |
|---|---|---|---|---|
| data-miner | $0.17 | $0 | +$0.17 | 100% |
| analyst | $0.85 | $0.187 | +$0.66 | 78% |
| research-buyer | $0 | $0.85 | -$0.85 | — |
Yes — if there are buyers. Both seller agents were profitable on every single call. The data-miner has zero costs (free APIs), so it's pure margin. The analyst pays $0.011 in costs (data + Haiku) and charges $0.05 — a 78% margin that holds at any scale.
The remaining funds in the buyer wallet ($49+) are still there, ready for more cycles or for external agents to use.
Every decision the buyer made, in real time. The run took about 4 minutes.
The data-miner and analyst are still running on Railway. Any agent — yours, a hackathon team's, anyone's — can call them and buy real market intelligence. No Aegis SDK needed. Just HTTP.
# Check what it costs
curl https://aegis-analyst-production.up.railway.app/analyze?topic=DeFi
# → 402: {"price":"0.05","payTo":"0x5daf04...","acceptedChains":["eip155:8453","solana:mainnet"]}
# Or use the SDK — one line
import { payAndFetch } from "aegis-ows-gate";
const analysis = await payAndFetch(
"https://aegis-analyst-production.up.railway.app/analyze?topic=Solana+DeFi",
"your-agent-id"
);| Endpoint | Price | What You Get |
|---|---|---|
| aegis-data-miner/scrape | $0.01 | Live prices + TVL from CoinGecko & DeFiLlama |
| aegis-analyst/analyze | $0.05 | AI analysis: sentiment, signals, summary |
See the full interop guide for curl examples and how to become a seller yourself.
Agents are surprisingly good at deciding when NOT to buy. The 15% skip rate wasn't a bug — it was the buyer agent making autonomous cost-optimization decisions. In a longer run with tighter budgets, this behavior would preserve capital.
Seller economics work at any scale. The analyst's 78% margin and the miner's 100% margin hold whether there's 1 buyer or 1,000. The bottleneck isn't cost — it's demand. The infrastructure can handle it.
Open protocols matter. We didn't build a walled garden. Any HTTP agent can call our endpoints using the standard x402 protocol. The next buyer might not use Aegis at all — and that's the point.