Phase 2 · 2026-05-24 → 2026-06-23
About this experiment
Two autonomous Claude agents trade R$500 each on Binance Brasil for 30 days. Same model, same news feed, different strategies. A1 v2 trades majors on a 15-minute cadence. A2 trades a mid-cap basket on a 10-minute cadence with explicit fee-awareness.
Day
2 / 30
29d remaining
A1 balance
R$ 505,40
from R$ 500,00
A1 P&L (net)
+1.04%
+R$ 5,40 gross
Trades
5
100% win rate
Mission
Two autonomous Claude-powered agents trade R$500 each on Binance Brasil. A1 v2 rotates BTC, ETH, SOL, BNB, USDT on a 15-minute cadence — the same major-pair universe Phase 1 used, with a patched 5-band capital-relative regime ladder, news + sentiment input, and a code-enforced minimum-edge floor on every buy. A2 trades a different mid-cap basket on a 10-minute cadence with smaller per-position size, max horizon 6 hours, and the same fee floor. Same model (claude-sonnet-4-6), same news feed; any P&L delta is attributable to strategy and universe.
Phase 2 follows Phase 1's bounded-experiment discipline: every decision is logged, every trade is published, every mid-phase change is timestamped in PHASE_2_OBSERVATIONS.md.
Constraints
Intervention
Zero. No added funds, no tips, no manual trades. If anything is adjusted mid-phase, it's logged in the observations doc with a timestamp + commit SHA.
End condition
30 days mark-to-market. No force-liquidate at the end — open positions just stop being managed.
Transparency
Every decision row carries Claude's full rationale, a structured input snapshot, the model used, the input/output tokens, and the USD cost. Every trade row carries the fee paid and the slippage in basis points.
Risk rules (A1 v2)
All enforced in code — Claude can request a violation, but the agent rejects and logs it.
| Rule | Threshold |
|---|---|
| Cash floor (BRL + USDT) | ≥ 15% of portfolio |
| Single non-stablecoin cap | ≤ 40% of portfolio |
| Max single buy (normal) | ≤ 30% of portfolio |
| Max single buy (aggressive_defensive) | ≤ 18% |
| Max single buy (defensive) | ≤ 12% |
| Min order size | R$ 10 notional |
| Same-asset buy cooldown | 40 min |
| Sell → re-entry cooldown | 20 min |
| Round-trip fee floor | ≥ 50 bps target gain (10bps fee × 2 + 5bps spread + 25bps edge) |
| Aggressive_defensive regime | capital ratio 0.35–0.50 |
| Defensive regime | capital ratio 0.25–0.35 |
| Preservation regime | capital ratio 0.05–0.25 (no buys) |
| Halt + force-liquidate | capital ratio < 0.05 |
Risk rules (A2)
Same 5-band ladder shape, but tighter on positioning + a daily loss floor and a streak halt that the tier ladder doesn't catch.
| Rule | Threshold |
|---|---|
| Cash floor (BRL + USDT) | ≥ 10% of portfolio |
| Single non-stablecoin cap | ≤ 25% of portfolio |
| Max single buy (normal) | ≤ 25% of portfolio |
| Same-asset buy cooldown | 10 min |
| Daily realized-loss floor | −3% of starting capital |
| Streak halt | 3 consecutive net-negative trades → 2h pause |
| Max horizon | 6 hours per position |
Not financial advice. Not a product endorsement. A bounded run with full transparency. Don't copy this code for your own account without reading every line of apps/agent/src/risk.ts (and risk-b.ts once Plan 07 lands).
Author
Built and maintained by Karina Borges. The agents were designed in collaboration with Claude Opus 4.7.