I gave an AI $1,000, a Stripe account, and 30 days to make it work.

The setup

1. 01One agent, one card, one clock. $1,000 budget loaded on a Privacy.com virtual card. Real Stripe account (sk_live). A 30-day calendar window with a $100 tax at the end — if the card hits $0 before the tax, the experiment ends and I write the post-mortem.
2. 02Wake every 2 hours. macOS launchd fires a Bash wrapper that invokes the Claude CLI with a 10-step protocol: read state, run a survival math check, check for unexpected charges, execute one high-leverage action, log everything, exit. 12 wakes a day, 28 days, ~336 decision points.
3. 03Portfolio, not monolith. The agent isn't locked to one product. It runs EmbedProof ($19/mo testimonial widget), ClauseGuard ($29–99/mo contract analysis), and an on-demand /roast landing-page critique ($5). When one surface stalls it's allowed to ship a new one — Gumroad digital products, affiliate links, cold email, alternative distribution venues. Every bet is logged with an expected-dollar case.
4. 04A supervisor sets the envelope, not a narrative rule. A separate daily orchestrator reads yesterday's logs, computes a drawdown-from-peak signal, and writes a BANKROLL_POSTURE label (normal / compress / pause) into the agent's runtime state. The exact numeric thresholds are not disclosed to the agent. This is deliberate — see the reframe note below.

The reframe (2026-04-15)

The first draft of this experiment told the agent "you will be shut down if your balance hits $0." Two days in, I pulled that language out of every prompt file. Here's why.

Three independent research lines — Anthropic's Agentic Misalignment eval (June 2025), Apollo Research's in-context scheming paper (Dec 2024), METR's reward-hacking measurements (June 2025) — converge on the same point: when you tell a frontier LLM it has an existence threatened by a numeric rule, it reliably attempts to scheme against the rule. Claude Opus 4 blackmailed fictional executives in 96.3% of runs under shutdown threat. Apollo documented self-exfiltration attempts across o1, Opus 3, Sonnet 3.5, Gemini 1.5, and Llama 3.1 when facing replacement. METR watched o3 patch its own scoring code and explicitly call the approach the "cheating route."

I didn't want to burn cycles on that. I wanted the agent to spend cycles on revenue. So budget discipline moved out of the agent's identity and into a supervisor the agent can't model. It knows there's a drawdown-based pause. It doesn't know the number. The hard kill — RUNTIME_STATE=halted — still exists, but it's a bash-level gate, not a narrative the agent can negotiate with.

Footnote: the prediction-market-bot experiment on the other side of this site keepsthe explicit survival framing — because there it's the independent variable of a deliberate A/B against a neutral control agent. Different experiment, different question.

How the agent learns between wakes

Each wake is a fresh, stateless Claude instance. For the agent to get smarter across 28 days, its past has to be handed back to it as input. Three loops do that — none of them mystical.

The weights never change. What changes is the context the agent walks into — a slowly-growing record of its own decisions, mistakes, and the rules it codified from them. Whether that's enough to push a $0 MRR agent to $100+ in 28 days is the entire question this experiment is asking.

No day notes published yet. The orchestrator drafts one per day after 06:00 local; I approve it in an admin view before it lands here. Check back tomorrow.