Agent Task

How it works in practice

A workflow template🎻Workflow TemplateWorkflows & AgentsA reusable workflow templateView reference → for onboarding a new integration defines five tasks. Three are human (sign the contract, approve the design, confirm go-live). Two are agent tasks: generate the API client from the OpenAPI spec and write the smoke tests. When the workflow runs⏯️Workflow RunWorkflows & AgentsAn execution of a workflowView reference →, the agent definition🦾Agent DefinitionWorkflows & AgentsAn autonomous agent definitionView reference → spawns the two agent tasks; each runs to completion and posts its output for a human to review at the next gate. The agent task carries the same fields as any task, a description, a status, an owner, so a mixed human-and-agent workflow stays one coherent plan rather than two parallel ones.

Agent task vs. its neighbours

• Task is the general unit of work, owned by whoever or whatever is assigned. An agent task is the subtype owned by an agent. Treating it as a sibling of the human task keeps a workflow legible when responsibility crosses the human-software line.
• Agent definition is the configured agent that performs the task. The definition is durable and reusable; the task is one bounded piece of work it is asked to complete.
• Workflow template is the plan that lays out which tasks run and in what order. The template defines the task slot; the agent fills it.

In the graph

In the Unified Product Graph, an agent task sits in the automation region alongside the other agent primitives and the ordinary task☑️TaskProduct SpecificationA unit of work within a story or epicView reference →. Two edges place it: workflow_template_defines_agent_taskWorkflow TemplatedefinesAgent Taskhierarchy slots it into a repeatable plan, and agent_definition_spawns_agent_taskAgent DefinitionspawnsAgent Taskhierarchy records which agent was handed the work. Because the agent task is a first-class entity rather than a flag on a generic task, a workflow that mixes human and autonomous steps stays queryable end to end, and you can ask which work an agent is accountable for without untangling it from the work people own.