Frequently asked questions

Functional AI is a runtime for agents. You build your prompt, agent, or multi-agent workflow as an Agentic Function; we host it, evaluate it, keep it running, and let you ship new versions anytime. You just call it from your code.

Anything from a single prompt to a multi-step agent or a multi-agent workflow. However you build it, it becomes one typed Agentic Function you call over an API.

An agent runtime is the hosted execution layer that runs AI agents in production: it executes the agent loop, evaluates changes before they ship, handles model fallback, and versions every change. Functional AI is an agent runtime. Full definition in What is an agent runtime?

Harness engineering is the discipline of building the scaffolding that turns a model into a dependable agent — tool interfaces, context management, prompts, guardrails, and evaluation loops. You engineer the harness; an agent runtime like Functional AI hosts and runs it in production. Full guide in What is harness engineering?

Any model. Use one we host with usage billed at cost, or plug in your own provider’s endpoint and credentials. You’re never locked into a single model or provider.

Functional AI monitors every call and automatically falls back to a backup model the moment your primary fails. Your Agentic Function keeps answering and your users never notice.

Define a dataset and a quality threshold, and every version is scored before it ships. Regressions fail the check, so you know a change is better — and cheaper — before a user ever sees it. See evaluating an LLM prompt before shipping for a full walkthrough.

Like any API. Promote a version and you get a stable, authenticated HTTP endpoint that returns typed output plus metadata — version, latency, cost, confidence, a status (success, partial, failed, or escalated), and a trace ID you can use to pull the full run. There's a Python SDK today (pip install fnai); more languages at launch.

Structured outputs — categories, enums, extracted fields — are graded by exact and field-level match, so the score is deterministic. Open-ended output (free text, agent transcripts) is scored with model-graded checks, which we treat as a signal, not ground truth: LLM judges carry known biases, so we track scores per version and expect you to keep a human review cadence as the calibration layer. Automated scoring scales coverage; it doesn’t replace judgment.

Defense in depth, and we’re honest that injection isn’t a solved problem. Tool output — web pages, documents, API responses — is treated as untrusted data, never as instructions. Tools run with least privilege, read-only access is separated from write access, and destructive actions (send, delete, pay, publish) can require human confirmation or run in a sandbox first. Secrets never appear in prompts, logs, or traces. The goal is to make injection harder and contain the blast radius if it lands.

It degrades instead of guessing. If a step, token, or time budget is hit, or a required tool is unavailable, the Agentic Function returns the best partial answer with an explicit partial or escalated status — never a confident wrong answer. From there you can route to a fallback or a human.

Yes. New versions go live behind the same function call, so you ship improvements without redeploying your app — and roll back instantly if you need to.

See the pricing page for launch plans. Usage on models we host is billed at cost; usage on your own provider is billed by that provider.

A step is any model call or tool call your Agentic Function makes. A model hop, a retrieval, a tool invocation — each one is a step. A single prompt is one step; a 12-step agent is twelve. Steps are the only thing we meter — never seats, never tokens. Eval steps cost a tenth of production steps, and model tokens pass through at cost. See the pricing page for the volumes included on each plan.

Yes. Functional AI treats all three as Agentic Functions. Start with a prompt and grow into an agent or a multi-agent workflow without changing how you call it.

Not yet — Functional AI is in private beta. Join the waitlist and we'll send an invite when it opens. Shipping AI in production today? Apply to be a design partner.