Status: accepted
Let an operator create a ready-to-run evaluation without first manually creating every reusable asset. A quick start is a declarative package that collects only the inputs it needs and then creates the complete evaluation configuration.
- From Builds → Create, choose Start with Quick Start or Configure manually.
- Select a quick-start card.
- Supply the manifest-defined parameters. Each field may provide a type, default value, description, options and placeholder.
- Review the input values before creation.
- Create the configuration. The operator can inspect or run the evaluation afterward; creation does not start a run.
The dashboard quick-start action opens this same flow directly.
A quick start has a qualified ID, semantic version, publisher, display name,
description, schema_version, parameters, assets and build template.
Parameters use lowercase identifiers and may be string, workspace, url,
select or a saved model_profile. A parameter may be required and may define
default, description, placeholder and options. Template values can
reference parameter values with ${parameter_key}.
The manifest declares these required assets:
- runner source and metadata;
- manager prompt template;
- fixed target-AI test-case set;
- execution environment;
- target environment; and
- build.
It may also declare a model-profile asset. When it does, the form collects the connection settings and creates a new reusable profile before the evaluation build is created. Secrets remain environment-variable names; secret values are never put in a manifest or sent through the API.
Before writing, Orbit validates required parameters, runner syntax, supported workspace paths, referenced assets and ID conflicts. All generated asset IDs are instance-scoped. If any write or validation after writing fails, Orbit restores the prior state and leaves no partial quick-start instance.
Importing a quick start only stores and validates its declaration. It does not execute runner source or start a process. Imported runner code is visible as a created asset before any evaluation is run.
Show a full generated-asset preview before creation and validate parameter types and option values server-side.
| Quick start | Purpose | Runner |
|---|---|---|
| User journey smoke test | Validate a named browser journey with operator-provided path, actions and success evidence. | Browser journey runner |
| Site exploration review | Explore safe same-site links and assess the rendered experience with evidence. | LangGraph site exploration runner |
| Agent self-improvement | Validate prompt or agent changes in a Git repository against a fixed browser journey and retain rollback-ready evidence. | Native improvement cycle |
| AI SLO and behavior drift monitor | Repeatedly assess structured AI quality, safety, latency, and cost evidence against a fixed baseline and retain supervised improvement decisions. | Evidence-gated probe cycle |