Can Claude Trade Crypto?
I gave an autonomous Claude agent real money on Binance Brasil and published every decision it made. I never traded manually, never topped it up, never fed it a tip. Two phases ran; both are finished. The agent lost money both times — and that turned out to be the least interesting thing about it.
Two phases, both closed
Each tick ran as a scheduled Lambda: pull balance, positions, market data and news → build the prompt → call Claude → get a structured decision with a thesis → validate it against code-enforced risk rules → execute → persist state.
17 → 24 Apr 2026
One agent, five symbols
−3.97%
R$203.24 over 7 days on Claude Opus 4.7, ticking every 15 minutes across a locked five-symbol universe. 708 decisions, 16 trades. The fix that mattered: threading the last 20 decisions back into the prompt lifted the agent's plan-reference rate from 12% to 94%.
24 May → 22 Jun 2026
Two agents, open universe
−1.83% / −1.25%
R$500 each, running side by side for 30 days — one conservative, one aggressive scalper — over an open universe of the most liquid BRL pairs. 6,141 decisions. Realized P&L before fees was actually positive for both; fees, slippage and open marks took it back.
The trust boundary is the whole design
The interesting engineering here isn't the trading — it's giving a model real spending power without giving it the keys to everything. Credentials are split so that no single compromised component can both decide and cover its tracks.
The agent holds exchange and model credentials — but no database
It can place an order and call Claude. It cannot read or rewrite the record of what it did, because it has no idea the database exists.
The API holds the database — but no exchange keys
It accepts writes from the agent over a bearer token, and the agent's identity comes from that token, never from the request body. It cannot place a trade.
The dashboard is read-only, and public
It holds nothing. Every decision, thesis and cost figure is visible to anyone while the experiment is running — which is what makes the record worth anything.
Every risk rule — position caps, cash floors, loss-streak halts, the fee-aware validator — lives in code, not in the prompt. Across 6,141 Phase 2 decisions there were zero capital-floor breaches. There are three independent kill switches: a halt flag the agent reads each tick, the scheduler itself, and self-halts triggered from inside the agent. Never trust a model to respect a limit that code can enforce.
What it actually taught me
These generalize past trading — they apply to any autonomous agent that runs on a schedule and spends something real.
The thinking cost 40× what the trading lost
Phase 2's combined trading result was −R$15.37. The Claude inference behind it cost about US$112 — roughly 60% of the entire book. For a high-frequency agent, the model bill, not the task outcome, is the dominant cost line.
A cheaper model traded better
Switching one agent from Sonnet to Haiku cut cost per decision by 67% — and it won 64% of its closed trades over the following 16 days, against 39% before. Small sample, but reaching for the biggest model by default is not diligence.
The edge was already in the data
High-confidence trades were net positive on both agents; medium-confidence ones were net negative. Removing just the medium-confidence exits flips one agent's P&L positive. The signal was logged from day one and never used.
The harness matters more than the model
The single highest-leverage fix across both phases was giving the agent structured memory of its own recent decisions. Most problems that look like reasoning failures are missing-context failures.
Metering is a schema decision, not a dashboard
Phase 1's honest answer about spend was “I don't know precisely.” Logging tokens, cost and model per decision made Phase 2 exact to the cent — and shaped the phase more than any trading behaviour did.
The failures were record failures, not money failures
None of Phase 2's operational misses touched capital. Each one broke the record instead: unmetered recaps, a scheduled job still firing after the phase ended, an API that quietly capped its own history. For an experiment whose product is the record, that's the failure mode to engineer against.
The full write-ups
Phase 1 — seven days, one agent
What the P&L doesn't tell you: the architecture, the bugs that almost broke it, and the moment the agent called out its own bias.
Read the write-up →Phase 2 — two agents, thirty days, and the meter
Two agents traded R$1,000 for a month and lost R$15.37. The thinking cost US$112.03. This one is about the meter, not the market.
Read the write-up →Not financial advice and not a trading strategy — the agent lost money in both phases. The reusable part is the architecture, not the trades it chose.