Agents

Building a Robinhood trading assistant on Claude Code: an MCP connector for market data, a strategy engine with hard-coded risk rules, and a design where unattended order placement is structurally impossible, not just discouraged.

What Claude Code actually is, how it differs from autocomplete-style AI tools, and the concepts — tools, permissions, subagents, skills, MCP, hooks — that make it useful for real engineering work instead of toy demos.

A practical, opinionated guide to designing memory for multi-agent LLM systems: topology, structural tenant isolation, the short-to-long maturation pipeline, retrieval and clamping, concurrency, and the four canonical failure modes.