Agentic loop

Catalysm bots are not set-and-forget. They run inside an agentic optimization loop on a weekly cadence that tunes the strategy, judge, and sizing rules the pipeline uses every 15 minutes.

Most bot parameters are not hardcoded forever. The values you see in the docs are the current configuration. Agents can change them each week after reviewing live performance.

What the agentic loop can change

The weekly loop can update most of the knobs that drive each 15-minute window:

• Indicator and module weights: how much each signal contributes to the score
• Gate thresholds: dead zone, confidence minimum, expected value minimum, spread limits
• Pattern and regime rules: which setups are blocked or allowed in which conditions
• Position sizing: grade multipliers, edge ramps, spread penalties, regime size cuts
• Risk limits: daily loss cap, maximum order size, counter-trend penalties
• Calibration: how predicted probability maps to historical win rate

What stays fixed between loops is the pipeline structure: the same stages run every window (strategy → historian → judge → execution). Agents tune configuration inside that structure; they do not redesign the bot from scratch each week.

Observe → Optimize → Deploy

flowchart LR
  A[Observe] --> B[Optimize]
  B --> C[Deploy]
  C --> A

1. Observe

Agents ingest live performance data: win rate, PnL, fill rates, slippage, and market conditions. Signals that never filled on-chain are excluded from performance analysis.

2. Optimize

Based on observed metrics, agents propose parameter changes: position sizing, gate thresholds, indicator weights, and risk limits. The weekly review reads historian output and ranks recommendations before approval.

3. Deploy

Approved parameters are deployed to the live bot. The cycle repeats weekly, so strategies adapt as markets shift.

Why weekly?

Crypto prediction markets move fast, but daily churn creates noise. A weekly cadence balances responsiveness with statistical significance: enough data to make informed changes without overfitting to a single session.

What agents do not do

Agents do not manually place individual trades. The bot executes its strategy autonomously; agents only tune the configuration that drives it.