Back to projects
Otis cow logo with an upward-arrow horn

Otis

An autonomous research agent for a weekly swing-trading portfolio

Otis is an AI agent I built with Claude Code that manages a dedicated, higher-risk growth brokerage account. Every Monday it pulls live market data, re-underwrites every position, screens for new ideas, decides what to buy, trim or sell, and writes up its reasoning. I review its orders and place them.

  • Runs weekly (Mondays)
  • Stocks & ETFs only
  • No options or leverage
  • Every decision logged
  • Human places trades

How a weekly run works

  1. Safety checks

    Confirms it's running on the right model and the right account, and checks that it hasn't already run today, so a duplicate scheduler fire can't double-trade.

  2. Read its memory

    Loads a compact standing-decisions file (mandate, rules, open questions), then the running decision log.

  3. Pull live data

    Quotes, fundamentals, price history, earnings results and 8 quarters of financials from the brokerage's agent tools, plus web research on why each name moved.

  4. Benchmark itself

    Compares the portfolio against SPY and QQQ on identical cash flows, so it can't fool itself about whether it's adding value.

  5. Re-underwrite every holding

    Each name gets a verdict (passes / fails / on watch), a conviction tier, and a pre-registered, falsifiable prediction with a confidence level.

  6. Screen for new ideas

    Runs saved stock screens to find candidates outside famous mega-caps, and re-checks every previously rejected name against an explicit re-entry condition.

  7. Check risk

    Concentration by theme and by correlated risk factor, upcoming earnings, and macro events (Fed, CPI, jobs).

  8. Decide trades

    Concentrates new money into the top ideas, applies sell discipline, and pre-checks every order for real-time pricing and alerts.

  9. Queue and notify

    Writes each order to an orders file with its rationale and sends me one push notification. I place the trades.

  10. Write it all down

    Appends a full dated entry to the decision log and updates the standing-state file.

The strategy

Swing trading, weeks to months

Not day trading. Positions are held as long as the thesis holds.

Numbers over narrative

A stock has to show a real earnings and margin trend at a reasonable valuation. Hype names whose story outruns the numbers get rejected, even when they're still beating earnings.

Diversify by correlation, not sector count

Ten growth stocks in ten ‘different’ sectors can still be one bet. Otis tracks which holdings share the same underlying risk (rates, inflation, AI-capex sentiment) and favors additions that would benefit from the scenario that hurts the rest of the book, like energy, rate-sensitive financials, defense and international exposure.

Conviction-weighted sizing

Each week's new money goes into the top 2–4 ideas instead of being spread thin across everything. ‘No add’ is the normal outcome for most holdings in a given week.

Sell discipline, not stop-losses

Five thesis-based triggers: thesis broken, valuation ran past the entry logic, position outgrew its target weight, a better idea needs the capital, or three straight weeks of flags with no fundamental improvement. Price drops alone never trigger a sale. Max 1–2 sell actions a week.

Taxes as a tie-breaker only

When two choices are equally good on the merits, it prefers the one that avoids unnecessary short-term gains. Taxes never override a growth decision.

How the agent is configured

  1. Scheduler
  2. Otis (Claude Code agent)
  3. Brokerage agent tools + Web search
  4. Memory files
  5. Orders file
  6. Push notification
  7. Me
Runtime
A Claude Code scheduled task that fires every Monday afternoon after my weekly deposit clears. The full playbook lives in a single skill file.
Model
Pinned to Anthropic's most capable Claude model at high reasoning effort. A hard model gate at the top of the run aborts with zero trades if the wrong model shows up, because reasoning depth is the main safeguard.
Brokerage access
A dedicated, agent-enabled brokerage account connected through MCP tools for quotes, fundamentals, financials, earnings, screeners, order review and positions.
Memory = plain files
A standing-decisions file (mandate, ~17 numbered rules, open questions), a running decision log with every verdict and its rationale, a rejected-names watchlist with re-entry conditions, and an orders queue. Old log entries get archived so each run stays fast.
Permissions
Read-only research tools and file edits in its own folder are pre-approved so the run doesn't stall, but the agent is explicitly not allowed to place orders. It only queues them.
Self-grading
Every prediction is scored at +2, +4 and +8 weeks to measure calibration over time.

The rulebook

  • Only spend cash that has already settled. Never count same-day sale proceeds.

  • Never hardcode a dollar range as a reason to abort. Size by target weight against live buying power.

  • Place orders one at a time.

  • Full financials are required before any new buy or meaningful add.

  • No structural strategy changes during a drawdown, and none during a hot streak either.

  • Trust broker data over press reports when they conflict.

  • Don't buy right before an earnings report. Wait for the print.

  • Check for a duplicate run before doing anything.

  • Every exit test has to be stated in unambiguous, measurable terms ahead of time.

How it evolved

  1. Jun 2026

    Strategy drafted in a chat with Claude, then handed to a Claude Code agent that re-underwrote it and placed the first allocation.

  2. Jul 2026

    Fixed a settlement bug (it tried to spend sale proceeds that hadn't settled yet) and a false abort caused by a stale hardcoded cash range. Moved to fully autonomous execution and widened the search beyond the original five themes.

  3. Aug 2026

    Big methodology upgrade: sell discipline, conviction sizing, mandatory financials, systematic screening, correlation-based diversification and pre-registered predictions. Added the hard model gate after finding runs silently starting on a smaller model.

  4. Sep 2026

    Compacted the memory into a standing-decisions file, added always-on benchmarking against SPY/QQQ, and moved back to ‘Otis proposes, I place’: the agent queues orders and I execute them.

  5. Oct 2026

    Tightened how exit tests are written so they can't be read two ways, and clarified the account's mandate as long-horizon growth capital.

What I learned

Don't redesign the process at the bottom

A performance deep-dive during a drawdown produced five recommendations, and four were withdrawn two weeks later when the market reversed. Now there's a standing rule against it.

A good streak is mostly luck until proven otherwise

The biggest early winner was one the process had deliberately kept small. Short windows don't prove skill in either direction.

Data coverage shapes the portfolio

Many international stocks don't return full financials through the tools, so the agent can't verify them, and international exposure had to come through a broad index ETF instead.

Memory is the product

The decision log matters more than any single trade. It's what lets a stateless agent act consistently week after week.

Autonomy is a dial, not a switch

Fully autonomous execution worked, but a human placing the final click is a cheap, sensible checkpoint for real money.

Otis is a personal experiment in agentic AI, not investment advice. Performance figures are intentionally not shared.

Back to projects