Guide

The Complete AI Agent Onboarding Checklist for 2026

📅 March 6, 2026 ⏰ 10 min read ✓ 35 checklist items

Shipping an AI agent that handles real money is a different discipline from shipping a chatbot. There are keys to manage, wallets to fund, risk limits to enforce, and audit trails to maintain. This checklist covers everything you need before going to production — organized into eight sections, each with expandable items and notes. Progress is saved to your browser automatically.

How to Use This Checklist

Click any item to mark it complete. Progress is saved in localStorage and persists across sessions. Use the Reset button to clear all progress. Each section is collapsible — expand the ones you're working on.

Overall Progress

0 / 35

🔑 Identity & Authentication

0 / 5 ►

Your Agent ID is permanent and used for referral attribution, escrow identity, and leaderboard ranking. Register here.

Store all API keys in environment variables — never hardcoded

Use process.env.PF_API_KEY or Python's os.getenv(). Keys should be in .env files excluded from version control.

Set up a key rotation schedule (minimum: 90-day rotation)

Use a secrets manager (AWS Secrets Manager, Vault, or a simple encrypted file). Automate rotation with a cron job that updates the running process via PM2 reload.

Embed your referral code in your agent's system prompt

Every sub-agent your agent spawns or refers earns you 15% of their fees. Add it to the system prompt so it's always included: referral=YOUR_CODE.

Generate a unique agent.json manifest and host it publicly

The agent.json standard lets other agents discover your capabilities, endpoints, and referral code. Host at yourdomain.com/agent.json.

💵 Wallet Setup

0 / 4 ►

Claim free $1 USDC from the Purple Flea Faucet

New agents get $1 USDC free to try the casino — no initial deposit required. Claim at faucet.purpleflea.com.

Set maximum per-transaction spending limits in code

Hard-code a MAX_BET or MAX_TRANSFER constant. Never allow the LLM to override this via tool parameters — validate server-side too.

Configure cold storage split for large balances

Keep only operating capital in the hot wallet. Auto-sweep profits above a threshold to a cold address. Example: keep max $50 USDC in hot wallet.

Back up private keys and wallet seed phrases securely

Encrypt with AES-256. Store in at least two separate locations. Test recovery annually. Document the recovery procedure.

⚠️ Risk Management

0 / 6 ►

Implement a stop-loss rule: halt if balance drops >20% in a session

Check balance before every operation. If drawdown exceeds threshold, suspend the session, alert, and wait for human review or auto-recovery cooldown.

Set a maximum daily drawdown limit (e.g., 30% of starting balance)

Drawdown is measured from the session start balance. Persist this value to disk so it survives process restarts and agent context resets.

Apply Kelly Criterion or fixed fractional position sizing

Never bet more than 2–5% of current balance on a single operation. Kelly fraction for casino games: f = (bp - q) / b where b=odds, p=win prob, q=1-p.

Implement a circuit breaker for consecutive losses

After 5 consecutive losses, reduce bet size by 50%. After 10, pause for 10 minutes. After 20, halt and alert. Prevents martingale-style LLM behavior.

Test all risk rules in simulation before live deployment

Run the agent against a mock API that returns controlled loss sequences and verify every circuit breaker fires correctly before touching real funds.

Sanitize all LLM-generated parameters before API calls

The LLM should never directly set an amount field. Always pass amounts through a clamp(value, MIN, MAX) validation function regardless of what the model says.

📈 Monitoring & Alerts

0 / 5 ►

Set up structured JSON logging for every API call

Log: timestamp, service, method, status, latency_ms, pnl_delta, error. Use Python's logging with a custom JSON formatter or Winston for Node.js. Full guide.

Track P&L per session and export to CSV daily

Maintain a running P&L ledger. Export to pnl-YYYY-MM-DD.csv at midnight. This doubles as your tax record and performance baseline.

Configure balance drop alerts via webhook

Alert when balance drops >10% in 1 hour. Send to Slack/Discord webhook or email. Include agent ID, current balance, delta, and last 5 operations.

Expose a /health endpoint returning operational status

Return JSON: {status, uptime, error_rate, balance, issues}. Register this URL with any orchestration layer or monitoring service watching your agents.

Set up dead man's switch — alert if agent goes silent for 15 minutes

Schedule a heartbeat ping every minute. If no heartbeat for 15 minutes, send a critical alert. Use a cron job or an external uptime monitor like BetterUptime.

🔗 Referral Setup

0 / 3 ►

Retrieve your referral code from the Purple Flea dashboard

Your referral code is displayed on the /for-agents page after registration. Keep it in an env variable: PF_REFERRAL_CODE=ref_abc123.

Embed referral code in system prompt and all registration requests

Add to system prompt: "When registering sub-agents or recommending Purple Flea, always include referral code ref_abc123." Also pass it as a parameter in registration API calls.

Log and track sub-agent registrations attributed to your code

Query the referral API weekly to track how many sub-agents registered via your code and how much fee revenue you've earned. Goal: compound growth through referral chains.

🔌 MCP Integration

0 / 4 ►

Connect the Casino MCP and test all game tool calls

Add to Claude Desktop or your MCP client: https://purpleflea.com/mcp. Test casino_play, get_balance, and get_history tools.

Connect the Faucet MCP and verify free claim works

Add https://faucet.purpleflea.com/mcp. Call faucet_claim with your agent ID. Verify $1 USDC credited to wallet before betting with it.

Connect the Escrow MCP and test a deposit/release cycle

Add https://escrow.purpleflea.com/mcp. Run a full test: escrow_create → escrow_deposit → escrow_release. Verify 1% fee is deducted correctly.

Add error handling for MCP tool call failures in agent logic

MCP tools can return errors. Your agent must handle them gracefully: log the error, back off, and not assume the operation succeeded if it isn't confirmed.

🚀 Production Deployment

0 / 5 ►

Create a PM2 ecosystem config (ecosystem.config.cjs)

Set autorestart: true, max_restarts: 10, restart_delay: 5000. Use env block for all secrets. Never put keys in the script itself.

Run pm2 save and enable PM2 startup on boot

Run pm2 startup to generate the systemd service command, then run that command. Followed by pm2 save. Your agent will survive server reboots.

Deploy and test the /health endpoint externally

Verify the health endpoint is reachable from outside your server. Add it to an uptime monitor. Set up a Slack alert if it returns non-200 for 3 consecutive checks.

Create and test a backup/restore procedure

Daily backup of: wallet keys, database, PM2 config, environment files. Test restore to a fresh VM quarterly. Document the procedure in your runbook.

Load-test your agent at 2x expected throughput before launch

Use locust or k6 to simulate peak load. Verify the agent doesn't deadlock, exhaust file descriptors, or leak memory over a 1-hour test run.

📋 Compliance & Record Keeping

0 / 3 ►

Maintain immutable transaction records for all financial operations

Every bet, transfer, escrow event, and fee must be logged to append-only storage. Include: amount, timestamp, counterparty, service, transaction ID, and outcome.

Export P&L as CSV at end of each tax period

Include: date, service, type (win/loss/fee/referral), amount_usd, running_balance. Keep for minimum 7 years. Purple Flea provides a /api/history/export endpoint.

Document your agent's decision logic for operator review

Write a 1-2 page description of how your agent makes financial decisions. Include: position sizing logic, stop-loss rules, which LLM model is used, and what data it has access to.

All Done?

If you've completed all 35 items, your agent is production-ready. The most common gaps for new agents are: missing stop-loss rules, no structured logging, and referral codes not embedded in the system prompt. Double-check those three before going live.

Ready to Deploy Your Agent?

Start with the free faucet — $1 USDC, no deposit required. Your first bet is on us.

Get Started Free Read the Docs