I gave an autotrader $360 and 30 days. I am not adding live money yet.

On May 14 I ran the kill-switch review on the live autotrader.

The decision is simple.

Keep V2 paper-only. Add no new live money. Revisit after the next scorecard.

That is not a dramatic kill. It is the boring version of discipline. The live book can keep being watched, but the next tranche does not go in just because I built the thing.

This is part of BMD HODL, the one-person AI-operated holding company I run nights and weekends. The cannon for this quarter says watch first, document everything, and decide from the rule instead of the sunk cost.

Today was the rule.

---

The setup

Two live accounts. Real money.

• Alpaca stocks: $200 deposited
• Kraken crypto: $160 deposited
• Total live: $360
• Compute: about $57 a month on Azure

The strategy is markdown-prompt-driven. Claude reads positions, market context, and a small playbook every morning. It proposes or manages trades inside guardrails.

Paper trading keeps running in parallel as the test bed. Any strategy change has to prove itself on paper before it touches live money.

That separation matters. Live money is where discipline gets tested. Paper is where experiments belong.

---

The numbers

Latest verified snapshot:

• Alpaca equity: $217.97 (+9.0% on $200)
• Kraken equity: $169.87 (+6.2% on $160)
• Combined: $387.84 (+7.7% on $360)
• Net of monthly compute: roughly minus $29
• vs SPY over the same window: minus 4.4%
• vs BTC over the same window: minus 10.2%

In isolation, +7.7% on $360 looks fine.

After compute, it is negative.

Against passive baselines, it is behind.

That is the whole point of the review. The bot does not get credit for being interesting. It has to beat the boring alternative or earn more time with better evidence.

---

The rule

The rule was written before the money went in.

If the live book is still net-negative after compute and still lagging both SPY and BTC, no new live tranche goes in.

If the live book is positive after compute or one benchmark has flipped, it can continue to the next tranche.

Today the rule says no new live money.

I am not routing the next $200 into the bot today. I am keeping V2 paper-only and waiting for the next scorecard.

---

What I am not doing

I am not declaring the system dead.

I am not pretending the result is good enough.

I am not changing the benchmark after the fact.

The live account did make money before compute. That matters. It also lost to the actual alternatives. That matters more.

The useful middle ground is to keep the live book contained, keep the paper system learning, and only promote capital when the scoreboard earns it.

---

Why this matters

Most builders are good at starting systems and bad at slowing them down.

Agents make that worse. Once a process runs on a schedule, it starts to feel alive. It produces logs. It writes reports. It gives you reasons to keep watching.

That is exactly why the rule has to exist before the result.

The rule is not there to punish the agent. It is there to protect the operator from narrative drift.

This autotrader is useful if it teaches me how to run capital with agents without getting high on my own software.

Today it taught the right lesson.

No new live money without better evidence.

---

The next scorecard

The next review is not vibes.

I want to see:

• Net result after compute
• Combined book versus SPY and BTC
• Paper V2 hit rate
• Cash drag
• Whether the strategy is learning from misses or just writing prettier reports

If those improve, I can add capital later.

If they do not, the live book stays capped and the next dollar goes somewhere boring.

That is not a failure. That is the system working.

---

If you are running agents near money, customers, or production, write the kill-switch before the run starts. That is what I built AgentGuard for. Budget caps, category caps, and breach hooks around agent loops.

Write the rule first. Code the rule next. Then let the result tell you what to do.