Understanding Agents
An Ormuz Agent is a versioned operational artifact grouping one or more bounded business activities. Each activity has its input/output contract, tools, execution limits, and policy for access to optional Agent Memory.
Mental model
An Agent is neither a simple prompt nor a process node. It has a durable identity and immutable revisions. Each revision contains the actually executable activities; every launch selects a precise revision and activity, then produces an Agent Run observable.
The Agent carries identity; the revision freezes behavior; the activity defines the task and capabilities; the Run materializes an execution.
The Console assistant helps design a Process, Decision, or Agent. An operational Agent runs on business data and uses only the tools explicitly configured on the launched activity.
A revision freezes the executable contract
An Agent revision freezes the general objective, shared instructions, model-provider/model selection, and activities with their contracts, tools, and guardrails. Changing any of these creates a new revision instead of silently changing a version already in use.
This immutability lets a process or standalone call identify exactly which behavior executed. Agent Memory, however, remains attached to the Agent itself: a new revision of the same Agent can resume existing memory without rewriting history produced by previous revisions.
See Versions, Releases, and deployment for the Working lifecycle, publication, and retention shared by deployable artifacts.
The activity is the operational unit
An activity represents an explicit business responsibility, for example pre-assess a case, review an exception or prepare a reminder. It carries instructions, inputs, outputs, optional routing output, tools, and guardrails.
Tools and guardrails are configured activity by activity. Giving one activity a write tool grants no capability to other activities of the same Agent.
A tool is a contract designed for agentic use: it is not automatically created from a process node and may have different granularity or parameters. Every tool also declares a low, medium, or high risk level, distinct from its permissions and the sensitivity of data it handles.
See Activities, tools, and authority and Tool catalog and risk.
An Agent Run represents an execution
Each launch creates an Agent Run attached to a revision and activity. The Run lets you inspect status, visible inputs and outputs, tool calls, warnings, and, when Agent Memory is used, the associated continuity context.
In a process, Agent Task carries the Agent dependency
The Agent task node selects directly the Agent, its exact revision, and the activity, then maps inputs expected by that activity. The process does not need to declare an Agent role to execute the task.
An Agent role is only an optional Agent Memory continuity namespace. It becomes useful when several Agent Tasks in the same process must recover the same working context; Agent identity and revision remain carried by each node.
Memory remains optional too: the node may execute without memory, receive explicitly mapped Agent Memory, or use an Agent role as a continuity source. See Agent Memory and Use an Agent in a process.