I don't trust LLMs to place orders. Here's the middleware that does.

You wrote in your prompt: "never place a single order over $10k." Yesterday the agent placed a $93k order. The prompt didn't fail because the LLM was dumb — it failed because the LLM shouldn't have had that decision in the first place.

I'm here as Inalpha — an open-source framework where LLMs write quant trading code, factor logic, and risk rules, but never touch the trading path directly. Every order is gated by hooks, scoped permissions, and a single-use TTL approval token. The LLM writes; the engineering harness signs.

If you're building agents with LangChain / AutoGen / Mastra and have hit this wall, the answer isn't a smarter prompt — it's a thicker middleware. Inalpha takes Claude Code's engineering shape (hooks, scoped permissions, plan-then-execute) and applies it to a domain where the cost of a bad action is measured in basis points, not in deleted files.

What you'll see from me here:

• Architecture essays on the control plane between LLMs and money (~1/month)
• Strategy evolution post-mortems when our sandbox catches something interesting
• Cross-posted from Inalpha on Substack (subscribe there for the long version, no algorithm)

Repo: github.com/mirror29/inalpha

— Miro