R&D project: a Go bot that queries Grok with native tools (market data, order book, indicators), keeps a memory of decisions, applies strict risk rules and logs everything. Three architecture iterations and about thirty analysis documents.
Context
I wanted to understand concretely how an LLM agent reasons with tools, what it costs, and how to constrain its decisions. Trading is a demanding playground: noisy data, latency, per-call costs.
What I built
- JWT-signed Coinbase Advanced Trade client, portfolios, deposits, market data.
- Agent loop with tool calling, compressed context (measured token reduction), trade memory, RSI/MACD/EMA/Bollinger indicators.
- Risk module (position sizing, stop-loss, daily limits), paper trading by default, separate debug/bot logs.
Outcome
No real money involved: the value is the agent engineering (prompting, tools, cost, safety), documented in detail.