Agent
An Agent is the primary autonomous entity in watsonx Orchestrate. It combines a large language model (LLM) with a set of tools, collaborators, and optional knowledge sources to carry out tasks on your behalf. When you send a message, the agent reasons about your request, selects the appropriate tools or collaborators, and works through the task step by step. Agents do not follow a fixed script — they reason at each step and adapt based on the results they receive. Agents come in three kinds:
Each agent carries a
name, a description, and a set of instructions. The description is especially important in multi-agent systems: supervisor agents use it to decide which collaborator to route a request to.
Assistant
An Assistant is a watsonx Assistant instance that you add as a first-class collaborator inside a watsonx Orchestrate agent network. You register an assistant by providing the watsonx Assistant service instance URL, the environment ID, and the authentication details. After registration, any native agent in your workspace can add the assistant to itscollaborators list, and the supervisor routes requests to it just as it does to any other agent collaborator.
Assistants support dialog and action-based automation patterns that differ from the LLM-driven reasoning of native agents. Use an assistant when you already have watsonx Assistant content that you want to preserve while embedding it in a broader, LLM-orchestrated solution.
Agent skill
An Agent Skill is a scoped, reusable instruction set for an agent. Where an agent’s top-levelinstructions field describes general behavior, a skill describes a specific procedure — processing an invoice, submitting an order, or running a data validation — that the agent loads only when that procedure is relevant.
A skill lives in a directory with a SKILL.md file at its root. The file has two parts: a YAML frontmatter block with name and description, and a Markdown body with the instructions the agent follows when the skill is active.
Tool
A Tool is a discrete, callable function that an agent uses to perform a specific action or retrieve information. Through tools, agents query databases, call external services, run Python scripts, and perform a wide range of other operations. Each tool has a well-defined input schema and output schema. The agent reads the tool’sname and description to decide when to call it, then passes the required inputs and receives a structured response.
watsonx Orchestrate supports several tool types:
An agent references individual tools by name in its
tools list. You can also group tools into a toolkit and reference them that way.
Toolkit
A Toolkit is a packaging construct that groups related tools so you can add or remove them from an agent in a single operation. It organizes tools by deployment model and lets you manage them as a unit throughout their lifecycle. watsonx Orchestrate supports two toolkit kinds:
When you import a toolkit, the platform exposes each tool inside it individually under the naming format
toolkit_name:tool_name. An agent references the toolkit by name in its toolkits list.
Agentic Workflow
An Agentic Workflow is a structured, graph-based orchestration of agents, tools, and human steps. Unlike an agent — which reasons freely at each turn — an agentic workflow follows an explicit sequence of nodes connected by edges, with defined branching, looping, and parallel execution. You import an agentic workflow into watsonx Orchestrate just like any other tool and attach it to an agent’stools list. When the agent calls it, the platform runs the full agentic workflow and returns the result.
Connection
A Connection holds the credentials that allow a tool, toolkit, or knowledge base to authenticate against an external system. Instead of embedding credentials inside a tool definition, you store them in a connection and associate the connection with the tool that needs them. Connections have a scope:member connections are per-user, while team connections are shared across all members of a workspace.
Knowledge Base
A Knowledge Base is a document collection that an agent queries to ground its responses in specific content. Where a tool fetches live data from a system, a knowledge base retrieves relevant text from documents that you upload or connect to. An agent attaches one or more knowledge bases via itsknowledge_base list. When your request matches content in the knowledge base, the agent receives the relevant passages as context before generating its response.