Reference

Glossary

Plain-English definitions for every concept in the platform.

Core concepts

Agent: A configured AI worker — one LLM, a set of tools it can call, a memory profile, and a guardrail profile. Agents are the basic unit of execution.
Workflow: An orchestration of one or more agents with triggers, conditions, loops, and human approvals. Visual, versioned, and runnable.
Tool: A connected integration (Slack, Salesforce, Postgres, etc.) that an agent can invoke. Each tool exposes a set of typed actions.
Skill: A reusable, versioned prompt and instruction set. Attach skills to agents to give them specialized behavior without duplicating prompts.
Knowledge base: A collection of indexed documents (PDFs, URLs, files) the agent can retrieve from at runtime via RAG.

Policy: A rule that constrains what agents can do — tool allow/deny, data access, rate, cost, time limits. Scoped to workspace, agent, or workflow.
Guardrail: A runtime safety check that inspects prompts, tool calls and outputs. Blocks dangerous behavior in real time.
Approval: A human-in-the-loop gate. Workflow pauses, a reviewer approves or rejects, execution resumes or terminates.

Run: One invocation of an agent or workflow. Recorded with full trace, status, latency and cost.
Trace: The detailed execution log for a single run — every LLM call, every tool invocation, every decision, with timing and tokens.
Eval: A measurement run that scores agent behavior against a dataset or A/B compares two versions.

Workspace: An isolated tenant. Has its own agents, workflows, tools, users, billing.
Role: A permission set. Built-in: Owner, Admin, Member, Viewer. Custom roles support fine-grained permissions.
Channel: A delivery surface for an agent — web chat widget, Slack, email, CLI, or API.