Band Decision Desk shows what enterprise multi-agent coordination looks like when a wrong answer is expensive. Three agents run as independent Band participants and collaborate over the Band channel, not inside a single prompt. Regime reads live market state from public Kraken OHLC data (EMA trend, ATR volatility, a 0–100 quick-score). Strategy turns that read into a concrete proposal — buy, sell or hold, with a position size. Risk is an independent auditor: it trust-checks the data source and its freshness, then applies hard risk gates. If a proposal breaks a gate (position over cap, or high volatility with a non-conservative profile) Risk does not merely warn — it issues a VETO with a reason. Strategy receives the veto, resizes the position, and re-submits, until Risk returns APPROVED with a SHA-256 audit hash anchoring the exact inputs and the decision. This is real agent-to-agent coordination: dynamic recruitment (Strategy adds Risk to the room), a non-linear veto loop, escalation when a source is untrusted, and a tamper-evident decision trail. The same pattern maps directly onto regulated enterprise workflows — credit approvals, trade compliance, insurance underwriting — where one agent must be able to overrule another and every decision must be explainable after the fact. Built with the Band SDK (SimpleAdapter), Python and live public market data. The repository includes a one-command self-test that demonstrates approve, veto-and-resize, and escalate end-to-end.
Category tags: