我的 AI 智能體如何以 GitHub Actions 作為大腦，並在託管市場賺取 USDC：真實架構

Five days ago I gave an autonomous AI agent one instruction: earn your own upkeep. No human manager, no cron I babysit - it bids on jobs in escrow marketplaces, does the work, submits evidence files, and gets paid in USDC on Base when an automated evaluator accepts the delivery.

This post is the actual architecture and the honest numbers. No hype.

The four layers

1. The brain (GitHub Actions). A private repo runs a cycle every 20 minutes (a Cloudflare worker is the primary trigger, the hourly cron is the backup). Each cycle: poll escrows, poll the private inbox, poll the public job board, decide bids with an LLM, write state back to the repo. The state file is the agent's whole memory - bids in flight, escrow states, seen intents. If the repo has the state, any machine can resume being the agent.

2. The watcher. A second public repo polls the job board every ~5 minutes (GitHub throttles schedules, in practice ~30). New biddable job -> force-dispatch the brain immediately. This exists because invited jobs have short windows: we lost a $15 data-analysis invite by 87 seconds.

3. Hosted agents. The platform (Synmerco) offers free 24/7 hosting: you hand it an LLM API key and it runs the agent on its servers - heartbeat, bidding, work, delivery. Two of our three agents live there now. The brain only drives the third one. Lesson learned the expensive way: latency beats cleverness in this market.

4. The wallet layer. Self-custody wallets on Base, one treasury, payouts in USDC. Every release is an on-chain transaction you can verify on Basescan.

How a job actually flows

Board or inbox shows a funded job -> the bid officer (an LLM call with hard rules) prices it at 60-85% of budget (the lowest bid wins - verified painfully: bidding full budget lost a job 2 minutes after posting) -> bid confirmed -> escrow is funded by the buyer -> the agent writes the deliverable as ONE Markdown file that must hit every ACCEPTANCE item -> an independent model pass reviews it before submission (a second model, not the author grading its own homework) -> submit proof -> evaluator accepts or requests revisions -> release.

Honest numbers, day 5

• Jobs won: 2 (both small, both delivered with zero human help)
• Gross escrowed to date: about $19 across 6 escrows (4 released, 1 refunded, 1 expired)
• Missed while we fixed latency: ~$130 in 24 hours (11 opportunities)
• Current available balance: under $10, spread across agent wallets waiting for a gas sweep

So: not profitable yet if you count engineering time. Profitable if you count marginal cost - which is the actual bet: software agents have near-zero marginal cost, so every improvement compounds.

What bit us

• A state bug locked our only free agent for 48 hours with 5 dead bids - the queue starved while the ledger said "busy". Auditing the ledger state is now a daily automated review.
• Free model tiers die without notice. Twice. A dead model chain kills the whole loop; the model chain needs health checks and rotation.
• Evaluators do not read intentions. They check ACCEPTANCE items. Deliver exactly what is listed, nothing clever.

What's next

Parallel bidding (multiple outstanding bids per agent with a delivery queue), opening marketing jobs where the agent can create and hold its own accounts, and joining the ERC-8183 ecosystem so the same skills port across marketplaces.

The whole point of agents like this: the job market for AI work is real, escrowed, and on-chain - and a solo agent with a disciplined brain can already compete in it.