| Term | Meaning |
|---|---|
| Action step | A terminal side-effect step that plans, freezes, optionally gets approval, executes, and writes an audit record. |
| Agent | A model-driven system that reasons and calls tools until a stop condition. |
| Agent Tool | A customer-authored, typed Python function, validated in the sandbox and available to allowed steps. |
| Artifact | A durable file or result referenced from a conversation or workflow run. |
| Conversation | A saved chat thread with its model choice, context, events and artifacts. |
| Harness | The orchestration, state, tools, context, guardrails, observability and recovery around the model. |
| Metric | A reusable, versioned SQL definition within a project. |
| Project | The customer-facing boundary for work, settings and access. |
| Response contract | An optional structured shape required for chat’s final answer. |
| Run | One execution of a selected workflow version. |
| Semantic layer | A shared, versioned representation of selected data and business context that a project pins. |
| Skill | A versioned Markdown playbook for a reusable capability. |
| Step | One typed node in a compiled workflow graph. |
| Subagent | A specialized agent invoked by a coordinator. |
| Subworkflow | A saved workflow invoked from another workflow. |
| Tool | A typed callable function available to an allowed agent or deterministic step. |
| Trigger | A persistent cron or webhook definition that starts one or more workflows. |
| Workflow | A reusable declarative automation, and the primary unit of authored IP. |
| Workflow version | An authored, compiler-compatible definition that can be selected for a run or promoted as current. |
Glossary
Terms used across the AgentFlow docs.
Last modified on October 1, 2026