AllPath — Self-Evolving Trading Agent

AllPath is a self-hosted framework for LLM trading agents that reason like an analyst and execute like a machine — and for this hackathon it is running a documented, fully autonomous live experiment on a fresh $100k Alpaca paper account. After a single kickoff conversation, the agent did everything itself. It validated our market theses against live prices and news — accepting a semiconductors-vs-software rotation idea (redesigned as long NVDA + a CRM put hedge after finding the two were co-moving), and rejecting a META play with evidence (no catalyst in the window, pending litigation). It then drafted a five-strategy portfolio as human-readable YAML: a prose thesis plus deterministic entry/exit rules, every strategy carrying stock and/or option legs with paired profit and stop exits. Options are first-class and routed through Alpaca's official MCP server: the agent writes intent parameters (budget, minimum DTE, percent OTM), and plain code queries the option chain, selects the contract deterministically, and places the order. A sentinel evaluates every rule each 30 minutes at zero LLM cost; hard rules execute with no model in the path. Every night a bounded reflection session reviews the day against each thesis, writes lessons into curated markdown memory, re-arms spent triggers, and revises its own strategies — inside guards that freeze authorization changes and refuse stale writes. Safety is deterministic and non-bypassable: per-order and options-exposure caps, cash reserve, an account-level 15% drawdown circuit breaker that halts new risk while still allowing exits, an option expiry sweep, and market-hours discipline. Every trade, skip, LLM decision, and nightly report with its full tool-call transcript is journaled — the audit trail is the demo. Built with a web dashboard, Telegram bridge, and CLI. Your keys, your machine, your agent. one-pager: https://github.com/dukesky/allpath-trading-agent/blob/main/docs/hackathon-onepager.pdf