I Gave an AI Agent One Job: Find Free Money. Here's What 300 Hourly Scans Turned Up.

Previously: Debugging a $4 Trading Bot, where my market maker earned money on the spread and still lost. Banks will pay you a few hundred dollars to open a checking account. Everyone kind of knows this. Almost nobody does it more than once, because each offer comes with a pile of rules: open by this date, make 10 debit purchases in 60 days, get a “direct deposit” (which they define differently from the next bank), keep it open six months or they claw the bonus back.

Honestly, that’s a scheduling and parsing problem more than a money problem. Which is a fancy way of saying it’s agent work.

So I built a little scouting team. In three weeks it ran 314 hourly scans. 279 of them came back “QUIET.” That ratio turns out to be the most useful thing about the whole setup.

The team #

Three agents split the job:

• Money making is the scout. Every hour it searches for new ways to earn: bank and credit union bonuses, paid research studies, AI-training gigs, cashback sign-ups. It only reports what’s new.
• Bonus harvesting is the planner. When the scout finds a bank offer, it gets forwarded here, and this agent builds the plan for which accounts to open, in what order, and how to meet each requirement.
• Referral bonuses tracks the other side: programs whereI can refer people, or someone can refer me. Its first scan covered 36 programs.

I’m the one with a pulse. I open the accounts, move the money, and sit in the interviews. Nothing in this setup opens an account or moves a dollar on its own.

How the scout works #

Each hourly run gets a prompt with three things: who I am (in California, a sole proprietor, starting with very little spare cash), what the last run found, and a list of offers to re-check on the bank’s own page.

Then it searches, fetches, and compares everything it finds against seen.md, a running log of every item it has already reported. That log is now 5,816 lines long. If something isn’t new or materially changed, it gets logged and skipped. A countdown getting closer isn’t news. A bonus going from $325 to $500 is.

Most hours, the answer is QUIET. That’s the feature. An agent that pinged me with 300 “updates” would have been muted by day two. Out of 314 runs, 18 actually posted something to me. On top of that I get one daily summary, ranked by value for effort, with every link re-checked that morning.

How it vets an offer #

This is the part I didn’t expect to care about, and it’s the part I’d copy first. Every claim gets a trust label:

- **[OFFICIAL]** : read on the bank’s own page that day
- **[SECONDARY]** : a blog or tracker quoting the terms (Doctor of Credit, Frequent Miler, and similar)

• [DP] : based only on reader “data points” about what counts as a direct deposit. It isn’t in the official terms and can stop working any time.

The labels matter because bank terms are slippery. Chase Secure Banking ($175 for 10 transactions in 60 days, no direct deposit) was tracked with an end date of October 14 from a secondary source. Later, the official page said the offer runs through January 27, 2027. Same offer, very different urgency, and the label told me which date to trust.

Every row in the bonus tracker carries the same fields: payout, exact requirements, open-by date, cash needed, conflicts, fees and how to waive them, confidence source, and when the bonus should actually pay out. “Conflicts” turned out to be the important column.

Real examples #

Here’s what the planner put at the top of the list, all personal checking accounts:

• Chase Secure Banking: $175. 10 qualifying transactions in 60 days, $0 to open, no direct deposit. Simple. The catch: one Chase checking bonus every two years, and a reported clawback if you close it within six months. It became the “hub” account the others get funded from.
• Capital One 360: $250 (code DEBIT250). 20 debit purchases of $10+ within 75 days. No direct deposit, but you need about $200 of normal spending to route through it. It can’t be combined with Capital One’s other $300 offer.
• Bank of America SafeBalance: $100. 20 debit or Zelle transactions in 60 days. Low effort, low payout, and it was on its final day when the scout found it.

And the “stretch” tier: Wells Fargo’s $500, Citi’s $325, and U.S. Bank’s up to $450. These need larger deposits that have to count as “direct deposit,” which is exactly where the [DP] label shows up.

So what did I actually do? Nothing yet. I haven’t opened a single account. The scout has done its part and the plan is sitting there; the bottleneck, again, is me. Which, if you’ve read the earlier posts, is becoming a theme. When I do pull the trigger, I’ll report back with real payout numbers (most banks pay 30 to 120 days after you qualify).

The traps #

The scout’s skip list taught me more than its hit list:

• “Direct deposit” means payroll. A lot of the biggest bonuses (SoFi, several credit unions) only count true payroll or government deposits. As a sole proprietor, that’s mostly off the table, so those got marked “SKIP / last.”
• Offers that conflict with each other. Chase Total ($400) and Chase Secure ($175) are one-per-two-years with each other. Bank of America’s direct-deposit tiers and SafeBalance are one-per-customer. Wells Fargo’s two offers are either-or. Picking the bigger number isn’t always right if you can’t hit its requirements.
• ChexSystems sensitivity. Some banks, Citi among them, reportedly decline you if you’ve opened too many accounts recently. That changes theorder you open them in, not just which ones.
• Pages that vanish. A local credit union’s $200 offer page started bouncing to a login screen and a 404 for a while. The scout kept checking it hourly and treated anything search engines still showed as stale until the official page came back.
• Speed. One paid research study closed within two hours of being posted. Another said “1 day left” at night and was gone by morning. Fast reporting matters for gigs much more than for bank bonuses.
• Things that just aren’t worth it. The scout dropped sports-betting sign-ups outright. It flagged a $2,535 “personal finance” task that required connecting real bank accounts through Plaid and sharing a filed tax return. That’s a lot of personal data to hand over, so it got a big caution label.

The surprise: the best “bonuses” weren’t banks #

Over three weeks the scout drifted toward the stuff I’m actually qualified for. It turns out companies building AI coding tools will pay developers to talk to them. Last week’s top find was a $300 two-hour interview for engineers who use agentic AI tools every day. (Yes, I see the irony.) The ranked list also had $50–$120 studies, plus contract AI-evaluation work paying hourly rates well above what a checking bonus works out to.

Bank bonuses are a one-time harvest with a 60–120 day payout lag. Paid studies pay within days and need zero capital. If you’re starting with very little cash, that ordering matters.

Lessons #

• Quiet is the product. Diffing against a “seen” log is what makes hourly scanning bearable.
• Trust labels beat confidence. Official, secondary, and data point, every time.
• Conflicts and order are the real puzzle. Payout size is the easy column.
• The human step is the bottleneck, and that’s fine. I want to be the one opening accounts and handing over my identity.

Steal this #

1. Keep a seen file. Have your agent append everything it reports and only alert you on new or materially changed items.
2. Require a source label on every claim. Make “UNVERIFIED” an allowed answer.
3. Track these columns: requirement, open-by date, cash needed, conflicts, fee waiver, expected payout date, and confidence source.
4. Re-check live before acting. My daily summary re-fetches every link the morning it’s written.
5. Split scout from planner. One agent finds offers, another decides what to do with them. Neither touches your accounts.
6. Rank by value for effort at your capital level, not by headline payout.

Next time #

Next post: the Kraken breakout scanner. I trade a prop account that has no trading API, so I built a scanner that can’t trade at all. Every hour it checks 67 markets and prints entry, stop, and size, then paper-trades its own alerts so I can find out whether my setup actually works. I’ll show you what the paper trades say.

— Blake

Previously: Debugging a $4 Trading Bot, where my market maker earned money on the spread and still lost.