Projects

A project is the main workspace for an AI product, assistant, agent, or workflow. Project navigation contains Improve, Behaviors, Monitor, Traces, Prompts, Tools, Evaluators, and Datasets. Use one project when the objects share production traffic, evaluation criteria, and deployment lifecycle. Create separate projects when teams need separate prompts, traces, datasets, or permissions.

​

What belongs in a project

• Prompts and prompt versions.
• Project tools and optional prompt MCP configuration used by those prompts.
• Evaluators and datasets.
• Production traces and monitor charts.
• Behavior analysis and Improve cycles.
• Deployment environments and history.

​

Project boundaries

Project boundaries matter because Platform workflows join data inside a project:

• Monitor and Traces read project traffic.
• Behaviors analyze project traces and subject-specific patterns.
• Improve cycles target prompts and behaviors inside the project.
• Evaluators and datasets are selected from the project context.
• Deployment environments are project-scoped release targets.

Use a separate project when traffic, deployment ownership, or evaluation policy should not mix. Use the same project when prompts, tools, datasets, evaluators, and traces are part of one product loop.

​

Project setup checklist

For a production project, create or verify:

• Prompts for each runtime behavior.
• Deployment environments for the release lanes your application uses.
• Provider credentials at the workspace level.
• API keys for runtime logging, deployment reads, and automation.
• Tools for external actions and data access.
• Evaluators for quality, safety, schema, cost, and latency requirements.
• Golden and regression datasets.
• Trace metadata standards for environment, route, release, and segment.
• Webhooks or cache refresh logic if your runtime caches deployments.

​

Common project patterns

​

Next steps