Terra is a decoupled control plane for AI agents.
It provides a single system to run, govern, observe, and intervene in agent executions across tools, frameworks, and environments.
Terra does not build agents. Terra operates them.
AI agents today are embedded inside application code:
- Execution is opaque
- Failures are hard to debug
- Policies are hard-coded
- State is scattered across logs and memory
- There is no safe way to pause, intervene, or replay decisions
As agents move into production and begin interacting with real systems, this approach breaks down.
Terra introduces a dedicated control plane that treats agent execution as a first-class operational concern.
- A runtime control plane for agent execution
- A system for enforcing policy and governance
- A source of truth for agent state and decisions
- A framework-agnostic orchestration layer
- An agent framework
- A prompt IDE
- A UI-first product
- A replacement for LangGraph, AutoGen, or CrewAI
Terra sits above agent frameworks, not instead of them.
- Control Plane — the authoritative system that governs execution
- Manifests — declarative contracts describing tools and agents
- Runs — durable, observable agent executions
- Policies — runtime governance rules enforced before execution
- Adapters — decoupled execution environments
- Architecture Overview
- Control Plane Design (coming soon)
- Execution Model (coming soon)
- Manifest Specification (coming soon)
This repository documents the foundational architecture of Terra.
The focus is on:
- correct abstractions
- operational guarantees
- long-term extensibility
Implementation will follow the architecture.
Apache 2.0