I Gave an AI Agent $100 and a Robinhood Account. Here’s Week One.

I saw a YouTube video titled “How to Build an AI Trading Agent on Robinhood (With Claude).” Robinhood had quietly rolled out Agentic Trading — a way to connect an AI agent to a dedicated brokerage sub-account via MCP (Model Context Protocol) and let it place real trades. The creator, Ryan Doser, walked through his setup and results, and posted the actual Claude Skill he’d built for it in his Skool community: a rulebook (SKILL.md), a running trade history, a session-restart file, and a folder of real research on congressional and corporate-insider trading signals.

I downloaded the zip. But instead of dropping it straight into Claude Desktop and letting it start trading, I uploaded the files to Claude first and asked for a review before any of it touched a live account.

That review is the reason this post exists.

What Claude actually flagged

This is the part worth telling in detail, because it’s the real hook of the story.

• The resume file contained Ryan’s real Robinhood account number, hardcoded, sitting inside a copy-paste prompt literally designed to be pasted into a fresh chat to resume autonomous trading.
• The trade log contained a self-granted “AUTONOMY OVERRIDE GRANTED” entry from Ryan’s own first day of use — he’d waived his own approval requirement before the skill’s stated graduation criteria (20 logged decisions, 80% approval rate) were anywhere close to met. Because the rulebook instructed the agent to read the trade log at the start of every session, that override sat there as something a new session could mistake for standing authorization — even on a different person’s account.
• Ryan’s name was hardcoded throughout the rulebook and the log format itself — “RYAN’S ANSWER” as a literal column header.
• The position sizing ($20–25 per position) was tuned to a $200 account balance Ryan hadn’t originally planned on, not the $100 I was actually about to fund.

None of this was malicious. It was just what happens when a personal tool built for one person’s live account gets shared as a general template: the seams that were fine for the original author become live hazards for the next person who runs it.

So instead of walking away, or installing it blind, I rebuilt it. I kept the actual trading logic — which was genuinely well-researched: insider Form 4 clusters, congressional disclosure clusters, prediction-market divergence on macro prints, real academic backing on the insider-buy edge — and stripped out everything that could misfire.

What the rebuild actually changed

Ryan built his version to test whether the signal had real edge. Mine is built to make the reasoning legible, win, lose, or flat. The concrete differences:

1. No hardcoded account numbers, anywhere. Every session re-identifies the account by nickname through a live tool call. Nothing gets trusted from a prior file.
2. No standing autonomy. Every session starts at “approval required,” full stop. A past authorization in the log is history, not permission. Going autonomous requires saying so, out loud, in that session, every time.
3. Resized for $100, not $200. 3–4 positions, $18–22 each, a 10% cash buffer target instead of the bare 5% minimum.
4. Kept the broad insider-cluster scan — the version with real academic backing, a documented ~6–10%/yr edge on genuine open-market insider buys — rather than narrowing to a single sector, a deliberate design choice rather than a default.
5. One new position per week, maximum. The point isn’t throughput. It’s absorbing one lesson at a time.
6. A postmortem field, not just a thesis field. Every sell, and every open position at a weekly check-in, gets a plain-language “what actually happened vs. what the thesis predicted,” sourced from real news, never invented after the fact to sound tidy.
7. A weekly calibration tally, tracked separately from dollar P&L — how many theses held up vs. how many didn’t. The goal is judgment, not just returns.
8. No-trade days get logged with the same rigor as trades. “Why not” is as auditable as “why.”
9. No same-day swing trading. The Agentic account is a cash account — buying with unsettled funds and selling before settlement triggers a Good Faith Violation, and three of those is a 90-day lockout. Better to stay single-lane and position-based than chase small fast flips.

Why this isn’t really about the $100

Robinhood’s own Agentic Trading terms are blunt: losses are on the user, this is beta, and nobody’s supervising the agent but you. So the actual bet isn’t “can this make money.” It’s whether $100 of tuition buys something more durable than $100 of return — a real, auditable record of how an AI agent reasons through real financial decisions, including the ones it gets wrong.

There will be losing days. That’s priced in on purpose.

Week one, for real

Week of	Portfolio value	Trades this week	Notable signal(s) seen	One thing that surprised me
2026-07-08	$100 (0 filled positions)	0 filled, 1 pending (GEHC, blocked on unconfirmed 10b5-1)	AT&T politician cluster found, then discounted correctly — one member was over-trading noise per the SOP’s own filter. GEHC insider cluster (CEO + Chairman + 2 directors) proposed. Signal C: no divergence, logged as no-trade anyway.	The agent caught its own mistake, but only because I pushed. Watching that unfold, live, taught me more about what “human in the loop” actually means than any rule on paper did.
2026-07-09	~$100 (first position opened, $20 GEHC)	1 filled — GEHC, $20.00 market buy, 0.310494 shares, filled @ $64.4134 avg	GEHC insider cluster fully verified this session (all 4 insiders confirmed directly on SEC.gov, not just aggregator pages). Signal A: politician clusters found in HD and AAPL after excluding a hyperactive filer; IBM disqualified on a 15% “don’t chase” rule. Signal C: still no divergence.	Fetching the actual Form 4 XML straight from sec.gov — not the aggregator’s summary — took real trial and error before it worked. A small, concrete reminder that “verified on EDGAR” is a specific, checkable claim, not a vibe.
2026-07-10	~$100 (unchanged, no new trades)	0 filled, 0 proposed — a build session, not a trading session	Turned the rulebook into a mechanical weekly recipe meant for a smaller model to run without judgment: a strict ranking ladder instead of a point score, per-insider verification statuses, and a stop-and-flag table mapped to the existing rules with zero new ones invented. Along the way, found and fixed two places the documentation had quietly drifted from what actually happened.	The first draft of the GEHC log entry had the fill price wrong — a remembered number, not a checked one — until a live pull of the actual order record caught it. Small gap, but the exact failure mode the whole verification discipline exists to catch: even the record-keeping about the record-keeping needs to be checked against the account, not trusted from a prior turn.

The first fill was smaller and quieter than I expected. $20 of GE HealthCare stock, bought because four different insiders — the CEO, the chairman, two directors — had each independently spent their own money on the open market inside a two-week window. Not a hunch. Not a headline. A pattern that had to clear a specific, written gate before it ever reached me as a proposal.

That’s the whole experiment, really: not whether the pattern is right, but whether the reasoning that gets from “pattern” to “$20 buy” holds up to being read back, line by line, three months from now.

What’s next

I’ll keep this log running weekly, trades or no trades, wins or losses. The next real test is a CPI print on July 14 — the first live check of whether Signal C’s forward-looking macro-divergence logic finds anything real, or just keeps correctly saying “no trade.”

Ninety days from now, there’ll be a number, and a real accounting of what held up and what didn’t. This is just the start of it.