Agent guide and operating notes

For the Crypto Agent this is usually the strategy working as designed: it decides once per day, and most days the right decision is to keep holding the same coin. It also deliberately refuses rotations that don't clear its quality bars (the hysteresis band and the breakout filter) — our backtests showed those refused trades were, on average, fee-burning mistakes. For the Market Agent, candidates must clear composite score, signal, setup, liquidity, spread, risk budget, and critic review before capital deploys.

Because every attempt to make it faster tested worse on real data. Four-times-daily refresh, intraday crash stops, and quick profit-taking were each backtested and each lost to the plain once-a-day design — crypto majors trend over weeks, and extra trading mostly buys whipsaw and fees. A few trades per month is the expected cadence; the daily decision stamp (visible in diagnostics) shows the engine is alive and choosing to hold.

The learning system needs real round-trip trades (buy and sell) to calibrate. In the first 5–10 days, the agent operates in a cold-start mode where its win rate is unknown, which adds conservative ghost pressure to the risk budget. After roughly 50 completed trades (typically 3–4 weeks), the learning dimensions have enough data to meaningfully adjust behavior: confidence calibration, hold duration bias, regime-specific setup selection, and counterfactual feedback all begin to produce actionable signals.

Multiple layers of protection exist. First, the deterministic scoring engine operates independently and can trade without the LLM. Second, the critic LLM independently reviews every proposed trade and can reduce size or reject. Third, deterministic gates enforce hard rules: regime-adaptive stop-loss floors, giveback profit-taking, risk budget caps, minimum edge thresholds, and correlation guards. Fourth, a circuit breaker monitors daily PnL and locks the sleeve if losses exceed thresholds. Fifth, the learning system feeds back outcomes so the PM sees its own track record. If the LLM is unavailable, the agent enters emergency safe mode where only protective exits are permitted.

Stop-loss exits are enforced by hard deterministic floors that scale with market regime: −1.85% to −2.20% for core assets and −2.50% to −3.00% for non-core. When a holding breaches its floor, the engine forces an exit regardless of what the LLM believes. Giveback exits can also trigger if a position gives back too much from its peak, even if still slightly profitable. The agent learns from these exits and adjusts future entry criteria for that setup type and ticker.

The critic layer can reduce any proposed trade size via a size multiplier (0.1x to 1.0x). Common reasons include correlation risk (too many similar positions), low conviction relative to size, or elevated regime transition risk. The improvement system can also apply a sizing reduction when it detects a correlation spike. The system tracks critic accuracy over time through counterfactuals, so if the critic is consistently wrong, the PM learns to push back.

The improvement system continuously tests configuration variants by replaying recent market cycles. It maintains a live challenger running in parallel and up to 3 shadow challengers running offline. When a challenger demonstrates a meaningful edge over the current champion configuration, it gets promoted automatically. If the promoted config underperforms, it rolls back. You can see the current state of the improvement system in the diagnostics console under the Decision Engine or Optimizer tabs.

The agent uses Coinbase's unrealized PnL as the source of truth for cost basis. After sell-and-rebuy cycles, Coinbase's data can carry stale cost basis from older trades. The agent includes a stale peak auto-heal: if the stored profit peak diverges more than 30% from the current gain, it resets automatically so giveback exits are never permanently blocked.