| Graph | Nodes (steps) and edges (control flow). One node is the entry where a run starts. |
| Nodes | Typically include LLM calls, branching, tool use, and outputs. |
| Tools | Named capabilities the model can invoke. Names and JSON Schemas must be registered on the service so the graph can reference them. |
| Inference providers | Agent nodes can run on multiple model backends, including OpenAI, Anthropic, Google, Groq, and Baseten. Depending on your deployment, plugin adapters can also route through AWS Bedrock or locally hosted models. |
| Validators | Optional output format rules (JSON Schema) attached to nodes to validate structured outputs. |
| Version | Published graphs have a version; API calls can pin agent_version or use the default active version. |