Platform & Architecture
You don’t need to know how the platform is built to use it. But the split between configuration and your data explains why each task is isolated and how an Enterprise deployment can keep your data in a dedicated environment. This page is the map.
The two planes
Everything in Introspection lives in one of two parts.
The Control Plane is operated by Introspection and holds identity and configuration: organizations, members, projects, deployments, API keys, integrations, and billing, plus the agent-config resources you author against (recipes, runtimes and runtime groups, experiments, judges, endpoints, variables, and write-only credentials).
The Data Plane holds your agent’s data: tasks, conversations, events, observations, and patterns, and files and shares.
Because your data lives in the Data Plane and not in the shared Control Plane, a Data Plane can operate on its own infrastructure, and on Enterprise plans in your own cloud account, for data-residency and compliance needs.
| Plane | What it holds |
|---|---|
| Control Plane | Orgs, members, projects, deployments, API keys, integrations, billing; recipes, runtimes, runtime groups, experiments, judges, endpoints, variables, credentials |
| Data Plane | Tasks, conversations, events / observations / patterns, files, shares |
What happens when a task runs
A single task moves through the platform like this:
- Create. An SDK call (or the channels gateway, or the dashboard) creates the task.
- Run. The task gets its own isolated sandbox: one environment for this task and no other. Your recipe is checked out at its pinned version and the agent loop runs.
- Observe. As the agent works, it produces the conversation: the replayable record of model calls, tool calls, and messages.
- Judge. When the task settles, the recipe’s judges grade the finished conversation.
- Teardown. The sandbox is destroyed. Any files the agent produced are saved durably to the Files API as artifacts. The conversation, its observations and patterns, and the judgements remain.
The record outlives the compute. The sandbox is disposable, the conversation is permanent.
Deployment and isolation
Your data is isolated by organization: every record is scoped to your org, and nothing crosses that boundary. How that isolation is provided depends on your deployment model: managed multi-tenant (shared infrastructure with logical isolation) or dedicated single-tenant (its own infrastructure). See Deployment for the comparison and what each means for data residency.
Introspection runs as a Kubernetes-native, GitOps-managed platform, which is what lets a dedicated deployment be stood up as its own isolated copy of the same stack. See Infrastructure for the components and security controls.
Related
- Core Concepts: the entities (recipes, runtimes, tasks, conversations) this architecture serves.
- Sandboxes: how a single task runs in isolation.
- Security model: tenant isolation, encryption, and how secrets are kept from the agent.
- Infrastructure & Data Handling: the operational controls behind the model.