Why Invariant?
Audience: Senior / Staff engineers building long-running agent systems. Not required: ML background.
Invariant is a Control Plane for AI Agencies. It solves the "State Management" problem in building complex, long-running agentic workflows.
What Invariant is NOT
Invariant is NOT:
- A prompt framework (like LangChain PromptTemplates).
- A chatbot engine (like Vercel AI SDK, though it can power one).
- An agent that "decides" its own execution loop (the Runtime decides).
The Problem
Most AI frameorks (LangChain, AutoGen) treat state as an in-memory side effect of the execution loop. This works for simple chatbots but fails for:
- Long-running processes (days/weeks)
- Human-in-the-loop approvals
- Auditability and debugging
- Deterministic replayability
The Invariant Solution
Invariant treats State as the Database.
- Pure Data Workflow: Workflows are defined in strict JSON, not code. This means they are portable, versionable, and language-agnostic.
- Centralized Control Plane: A single source of truth manages the lifecycle of every agent.
- Determinism: Given the same history and input, the Runtime Engine always calculates the same next step.
- Observability: Every state transition is an event in the database. You can replay, rewind, and fork executions.
Core Philosophy
- Invariants over Heuristics: Enforce strict rules (Type Safety, State Integrity) rather than relying on "fuzzy" LLM logic for control flow.
- Explicit Dispatch: The LLM decides what to do, but the Runtime decides when and how to execute it.
- Headless by Design: The Control Plane has no UI. It drives any frontend via a standardized State API.