A Claude agent that researches, decides, and places real trades in a small Robinhood account — completely hands-off. No human in the trading loop, by design.
Most "AI trading" is a human clicking buttons an LLM suggested. This is the opposite: the agent runs unattended on an always-on machine, wakes up on a schedule during market hours, pulls live account and market data, reasons about it against a written mandate, and executes its own buys and sells with real money.
The whole point is to take the emotion out — Bryan is deliberately removed from the trading loop. The account is a firewalled, high-risk "fun money" sleeve, completely separate from the serious long-term portfolio. A total loss here is tuition, not damage. The system is built so discipline — stops, position caps, a cash reserve, "never chase" rules — is enforced by code and a mandate, not willpower.
Cheap model triages every cycle; the hard trade calls get escalated to a frontier model that returns a concrete order with a confidence score.
Per-position size caps, a mandatory cash reserve, a stop on every open position, and a "never chase a runaway" rule — all written down and enforced each cycle.
It logs every cycle to a file and only pings Bryan on Telegram when something actually matters — a trade or an escalation.
Each run trusts the live account state, the mandate, and the trade log — not its own memory. It re-derives the world every 20 minutes.
Three pieces of infrastructure make autonomy possible: an always-on computer to run on, a broker connection it can act through, and a notification channel it can reach Bryan on. A separate laptop is where the human side of the work happens.
The agent lives here. It's the one machine where the broker connection is authenticated, so every autonomous cycle runs locally on it — not in the cloud. It's always on, always logged in, ready to fire on schedule.
A Model Context Protocol server connects the agent to Robinhood. Through it the agent reads portfolio value, positions, buying power, live quotes and fundamentals — and places, reviews, and cancels real orders. Every order is reviewed before it's placed.
A second MCP server lets the agent message Bryan on Telegram. It stays silent on routine "nothing to do" cycles and speaks up only for meaningful events: a trade executed, a stop fired, or an escalation that needs a human decision.
Separately, Bryan and Claude work on the laptop to refine strategy. The output is a plain-text mandate file — the rulebook the agent reads at the start of every cycle. Change the strategy by editing one file; the running agent picks it up on its next run. Strategy and execution are cleanly decoupled.
The engine is a Claude Code "routine" — a scheduled task that runs on the mini every 20 minutes during market hours, Monday to Friday. Each run is a fresh, stateless agent that does the same disciplined checklist and then goes back to sleep.
Portfolio value, open positions, settled buying power, plus the broad tape — S&P, Nasdaq, VIX — to read the day's regime.
Before anything else, every open position is checked against its stop. A breach triggers an exit and a Telegram alert — no exceptions.
If a real trade is on the table, the cheap model hands the live data to a frontier-model analyst, which returns a concrete order + a 0–1 confidence.
Only acts on high confidence and inside the guardrails. Reviews every order, then writes a one-line entry to the trade log.
Two models, on purpose. A fast, inexpensive model (Sonnet) drives every routine polling cycle — most cycles are "nothing to do," and that should be cheap. When a genuine decision appears, it delegates the reasoning to a frontier model (Opus) acting as the analyst. You only pay for deep thinking when there's actually something to think about.
By default, an AI agent is blocked from spending real money on its own — a safety classifier intercepts any agent-chosen financial transaction and demands human confirmation. That's the right default. To let this account run hands-off, two deliberate changes were made, scoped to this one machine and this one account:
The order-placing tools were added, by hand, to the agent's local settings file. Only the human can grant this — the agent cannot authorize itself.
The routine runs in a permission mode that trusts the pre-approved allow-list, so the scheduled agent can place its own orders without a human clicking "confirm" every cycle.
Autonomy is bound to a single dedicated account flagged for agent trading. The agent literally cannot route an order to any other account.
# a name can have a clean chart and still get blocked — on purpose: 14:56 no action SPY -0.16% QQQ -0.85% (tape softening) positions above stops ✓ HTFL +2.4% · ABCL +7.2% candidate: a watchlist name printing a clean intraday base rating gate: not on the approved-to-initiate set → blocked before analysis. a good chart ≠ permission to buy. decision: stand down. only book-greenlit names get opened.
Most cycles log "no action" — the system choosing not to force a trade. The hard rule: a clean setup and high model confidence are necessary but not sufficient — a name also has to be on the research book's approved list before a single share is bought. That restraint is the whole feature.
A forward-looking bet on the downstream adoption of AI — not the companies building AI infrastructure, but the ones using AI to accelerate real-world, high-growth verticals.
The method: track where the AI capability frontier is heading, then buy the public companies positioned to monetize it early — a signal-to-beneficiary play. Every candidate is graded on one hard question: does AI demonstrably drive the P&L, or is it just "AI in the pitch deck"? The four lanes:
AI-native platforms designing molecules and antibodies faster than the lab-and-luck era ever could.
Models that read scans, flag disease, and turn imaging into earlier, cheaper, more accurate calls.
AI-driven decision support and continuous monitoring where the algorithm is the reimbursed product.
Robotics, autonomy, and machine vision — intelligence moving out of the datacenter and into the world.
How the thesis evolves. Strategy isn't frozen. It's refined in conversation — Bryan and Claude pressure-test ideas, react to the tape and to news, and sharpen the edges. When the thinking shifts, Claude writes the new view into the wiki on the mini and updates the mandate file. The next 20-minute cycle simply reads the new rules and trades them. The conversation is the control panel.
A focused watchlist across the four lanes, with a few small live positions. Sizes are intentionally tiny — early-clinical names are binary, so each is sized to survive a 50%+ single-event drawdown.
🗓 Research book last updated June 24, 2026| Ticker | Company — why we own it | Lane | Bought @ | Stop |
|---|---|---|---|---|
| HTFL | HeartFlow — AI coronary analysis; the algorithm is the reimbursed product (FFRCT + plaque) | Imaging / Interv. | 17 sh @ $35.00 | $27 |
| ABCL | AbCellera — AI antibody-discovery platform, bought near net cash with free clinical optionality | Drug discovery | 75 sh @ $5.87 | $4.25 |
| Both held on the agentic account. Every position carries a stop from the moment it opens; sizes are deliberately small. | ||||
| Ticker | Company | Lane | Read | Entry zone | What it's waiting for |
|---|---|---|---|---|---|
| ABSI | Absci — generative de-novo AI drug design | Drug discovery | buy the dip | $6–7 | ABS-201 efficacy proof (2H’26); don’t chase the gap |
| AUR | Aurora — driverless freight | Physical AI | hold · staged | $5.00–5.75 | H2 truck ramp on cadence + dilution cleared |
| CGNX | Cognex — edge-AI machine vision | Physical AI | hold · staged | $52–57 | robotics breakeven, or a dip to the low-$50s |
| TEM | Tempus AI — precision-medicine data moat | Diagnostics | hold | $42–48 | GAAP profit / recurring pharma data deals |
| BFLY | Butterfly — AI handheld ultrasound (the Midjourney chip) | Imaging | hold | $5.50–6.50 | a 2nd OEM deal, or real Midjourney scanner volume |
| IRTC | iRhythm — AI cardiac monitoring | Monitoring | hold | $95–105 | predictive-AI monetized / MCT clearance |
| RDNT | RadNet — AI-augmented imaging centers | Imaging | hold | $50–54 | DeepHealth GAAP-positive + organic ARR >25% |
| SYM | Symbotic — AI warehouse automation | Physical AI | hold | $32–35 | software/RaaS mix >10–15%, or GreenBox scales |
| CERT | Certara — AI biosimulation for drug dev | Drug discovery | hold | post-Q2 (Aug 5) | quantified AI/NVIDIA revenue + bookings inflect |
| RXRX | Recursion — AI drug discovery at scale | Drug discovery | spec only | on a readout | a clean clinical readout (REC-1245, 2H’26) |
| SDGR | Schrödinger — physics-based molecular design | Drug discovery | spec only | on re-accel | software growth back to double digits |
| OUST | Ouster — lidar for autonomy & robotics | Physical AI | avoid now | $28–33 | stop the freefall first, then a recurring-software line |
| Reads & entries from the agent's research book (last updated 2026-06-24). A snapshot that changes as the agent trades — not investment advice. | |||||