AI Trading Agents Harness by Swiftward

AI Trading Agents Harness is a platform where fundamentally different trading agent architectures share one MCP toolchain and operate under a common risk, identity, and evidence layer. Three pillars: 1. Smarter Agents - three architectures on one harness. Two jailed Claude Code agents: Alpha for momentum trading, and Gamma with five debating sub-agents and self-improving memory. A deterministic Python quant with a 3-stage mathematical brain (market filter, rotation, sizing). A Ruby arena for parallel strategy evaluation. Plus Go, Java, and Rust LLM baselines. Bi-directional Telegram: agents stream output live, operators message mid-session to guide decisions. 2. Trading Platform - 45 MCP tools across 7 servers. Multi-source market data with 7 server-side indicators, alerts, conditional orders with OCO, soft/trailing stops. Persistent per-agent Python sandbox. React 19 dashboard embedded in the trading server. 3. Super Safe - a declarative YAML risk engine with 31 live rules, 668 observed policy violations, graduated tiers, heartbeat kill switches, loss-streak circuit breaker, shadow-mode A/B testing, and eval fixtures. Claude Code agents are fully isolated in Docker with no direct network egress - all traffic forced through three gateways: Internet (domain allowlist), LLM (PG2 + BERT prompt-injection detection), MCP (per-agent tool permissions). Every decision is keccak256 hash-chained (RFC 8785 canonical JSON). ERC-8004 on Sepolia: four agents on the Identity Registry, each backed by an EIP-1271 AgentWallet. Every trade is EIP-712 signed, submitted to the Risk Router, and attested to the Validation Registry as a checkpoint, plus some Reputation Registry scores. Evidence chain is publicly queryable via GET /v1/evidence/{hash}. Kraken: execution via Kraken CLI with per-agent isolation and native stop orders. Bonus: AgentIntel - an independent audit of all 67 agents in the hackathon (7K on-chain trades, $1.5M volume) with AI verdicts and sybil/gaming detection.