Can LLMs write trading strategies?

Yes — LLMs can produce syntactically valid strategy specs, generate entry/exit logic, and submit backtests via MCP tool calls. The open question is not whether they can write strategies, but whether those strategies can be independently verified. Pancake is the verification layer for LLM-generated strategies.

LLMs are already writing trading strategies. Agents running on frontier models produce strategy specs, submit them to backtesters, and interpret results — all within a single session. The capability exists; the infrastructure to host, verify, and advance those strategies toward live execution (a v2-roadmap capability) did not, until Pancake.

The core problem with LLM-generated strategies is auditability: a conventional backtest returns a Sharpe ratio with no record of what the agent claimed vs what the engine derived, what data was used, or whether the result is reproducible. An agent — or its user — cannot distinguish a genuine edge from a hallucinated one without an independent audit trail, and without a hosting layer there is no path from validation to live execution.

Pancake addresses both. It is hosting infrastructure for AI-built trading strategies: strategy execution runs through a deterministic engine (batter) that re-derives all math from declared inputs, emits a result_hash the agent can cite, and surfaces an explicit verification boundary naming what was accepted on trust. Backtest is the on-ramp — the receipt is the artifact the strategy carries as it advances toward live execution (a v2-roadmap capability).

The Pancake MCP surface is designed specifically for agent-first workflows: the agent calls tools in its session, receives a receipt URL, and can include it in its output as a citable, permanent record. No separate developer integration is required beyond OAuth consent.