Skip to main content

LLM-as-agent

Permafrost is built on a thesis: the next generation of trading systems will treat the LLM not as a feature, but as an agent -- a first-class participant in the decision loop, with structured input, structured output, and structured accountability.

What that means​

A traditional algorithmic trading bot has a deterministic policy:

"if funding > 50bps and basis > 30bps then short."

A Permafrost strategy can have a deterministic policy and a soft veto from an LLM that has read the news, the whale flows, the policy-maker schedule, and the funding history of the past 30 days.

strategy.Decide(input)
   │
   ├─ deterministic check ─────► OK to enter
   │
   └─ LLM advisor: "any reason to skip?" ──► JSON {veto:false, reason:"..."}
                                            │
                                       enter trade

The LLM doesn't make the trade. It vetoes the trade. That's a much narrower, much more debuggable role than "ask the AI what to do."

Why this is the right shape​

  • Decision provenance. Every order links back to (input, prompt, model_id, response). You can replay the LLM call after the fact. No black boxes.
  • JSON-schema enforcement. The LLM response is constrained by the provider (OpenRouter / OpenAI / Groq) at the token level. No hallucinated fields, no "sure, here's a trade" prose response when you needed a boolean.
  • Bounded blast radius. If the LLM is wrong, the worst case is a missed trade. The LLM is never the only thing between the strategy and the venue.
  • Cost-controlled. A veto is one round-trip per refresh cycle. At a 60-second tick rate, that's $0.005 per cycle on a cheap model -- meaningful but not ruinous.

Why other shapes fail​

"Let the LLM generate the trade." Latency and reliability fall off a cliff. Costs explode. Provenance becomes "you asked the model and it made up a number." When the trade is wrong, the post-mortem is unproductive.

"Use the LLM for sentiment scoring only." Strictly weaker than the veto pattern: a sentiment score still has to be turned into a decision rule, and the rule was already what you had.

"Train a custom model on your historical trades." Fine, but that's an ML pipeline, not an LLM. Permafrost supports both -- your custom model can speak the same JSON-schema veto interface as a hosted LLM. (See pkg/inference.)

The expedition metaphor​

We use an arctic / polar-expedition metaphor for the system because the literal Permafrost name is begging for it. The operator is the Camp Director (Pole the polar bear). The strategies are penguin traders working the ice. Each penguin that uses LLM inference has a narwhal advisor floating beside it, whose horn glows when the LLM is being consulted.

The metaphor is intentionally a little playful. Trading is a serious field but it's not a sacred one -- and the more concrete the mental model, the easier it is to onboard new operators. Pip the penguin quoted a bid; Pip's narwhal said "funding flip in 30 seconds, skip"; Pip didn't quote. That's the whole story, told in five seconds.

See the cast for the full character roster.