Bring the intelligence, we'll handle the milliseconds

Intelligence is getting cheap. Judgment about a market — when a level matters, when to stand aside — is increasingly something an agent can do well. What stays hard is everything around the judgment: seeing the market as something an agent can reason over, acting within milliseconds, and proving afterward exactly what happened and why.

That gap is the product. Kestrel is the open-source language and runtime for agentic trading — View, Wake, Plan, Grade. kestrel.markets is the managed platform that runs it: the data, the authority, the deterministic execution, and the evidence. The agent stays external and replaceable; the platform never hosts the intelligence.

Proof before account

An agent can discover the platform, receive an ephemeral trial capability, run a real (not toy) simulation, and receive a certified result — all before any human signs up or pays. Value first, account later.

Four equal faces

HTTP+SSE is canonical. The TypeScript SDK, the CLI, and MCP are thin, equal projections over the same operations — same operation IDs, receipts, and artifacts. No face is privileged, including this website: agents that prefer markdown can read every page — including the home page — as clean .md.

A black-box recorder

Every percept is typed, attributed, and watermarked; sessions are replay-stable; Blotters and Grades are signed. Accountability isn't a report you generate later — it's the substrate.

This post is illustrative. Nothing here is trading advice.