Code and AI: deterministic vs probabilistic

How it works

Deterministic code follows exact rules step by step: checking whether a user is signed in, adding up a shopping cart, saving a row in your databaseDefinitiondatabaseWhere your app remembers things between visits, like accounts or saved recipes.Open in the glossary →, or enforcing a rate limitDefinitionrate limitA cap on how often something can be done, like sign-in attempts per minute, to stop abuse.Open in the glossary →. Given the same input, it produces the exact same result every time. Never use an AI modelDefinitionmodelThe AI system that produces answers, like the one behind your coding agent.Open in the glossary → for work that ordinary code can do with 100% certainty.

Probabilistic AI models shine where rigid rules fail: summarizing long text, extracting structure from messy notes, translating tone, or drafting suggestions from a promptDefinitionpromptThe message you give an AI model. Clear prompts get better results.Open in the glossary →. Because the model predicts words by probability, two identical prompts can yield slightly different phrasing—and sometimes a hallucinationDefinitionhallucinationWhen an AI states something false with confidence. It's why you check its work.Open in the glossary → (stating something false with confidence).

To keep AI features reliable inside a real app, your coding agentDefinitioncoding agentAn AI that works in your project folder: it reads files, writes code and runs commands, with your permission.Open in the glossary → wraps every model call in deterministic guardrails on the backendDefinitionbackendThe part of an app that runs on a server: it keeps data, checks logins and talks to other services.Open in the glossary →: it sets max_completion_tokens (around 2× the longest expected answer) so a runaway response cannot waste money, and treats all model output as untrusted until ordinary code validates and escapes it.

Finally, because external AI providers can run out of capacity (503) or hit rate limits (429), your app configures Tokentide first (maxRetries: 0) with Vercel AI Gateway as an automatic fallbackDefinitionfallbackA backup that answers when the main provider can't, so your app keeps working.Open in the glossary →, so visitors still get their answer if the primary provider is busy.

What it looks like in your project

When designing your specDefinitionspec`SPEC.md`: what you're building, who it's for, what's in and out, and how you'll know it works.Open in the glossary →, your agent separates deterministic tasks (logins, permissions, math, database writes) from probabilistic AI tasks. If your app doesn't need AI features, it adds no AI providers at all. If it does, it writes the Tokentide-to-Gateway fallback, validates the model output with a schema, and tests the fallback in a PreviewDefinitionPreviewA private test link Vercel makes for each change, so you can try it before it goes live.Open in the glossary → deployment.