What an agent is made of
An agent also has a name, an optional description, and an id.
How a run moves
An entry runs its steps, if it has any, then starts the sub-agent it names. From there, every sub-agent names what comes next: another sub-agent, or the end of the run. The dashboard’s starting arrangements are patterns of these links. A single agent starts with one sub-agent that you can expand later. A sequential pipeline chains sub-agents, and each one receives the reply of the one before it. A supervisor routes work to specialists by calling them as tools. A parallel fan-out hands work to several workers and collects their replies in an aggregator. A run has no separate response step. Its answer is what its sub-agents said, and the entry decides who receives it.Sub-agents
A sub-agent is one model with one job. Its provider is a stored connection to an AI service, and the model belongs to the provider. Two sub-agents that need different models use two providers. On ServFlow Cloud, a sub-agent with no provider uses built-in model access. The system prompt sets the sub-agent’s role. The user prompt is a template filled in on every run from what the entry delivered, such as{{ body "question" }} for a webhook or {{ .discord.content }} for a Discord message.
A sub-agent can call tools while it works: an action, tools from an outside service, another sub-agent, or another agent. It can also have a workspace, a folder of files it reads and writes with built-in file tools. A workspace is how a sub-agent keeps notes from one run to the next.
Steps
A step is an action an entry runs before its sub-agent, every time, to fetch what the prompt needs or to do something first. Each entry has its own steps, because what is worth fetching depends on what arrived. See Steps before the agent for how they work and when to use a step instead of a tool.Identity and state
Every agent has an id made from its name, such asquestion-answerer. Anything that refers to the agent uses that id, including a tool on another agent that hands it a task.
An agent is either enabled or disabled. A disabled agent stays stored, but nothing can start it until you enable it again.
Creating an agent in the dashboard stores it disabled. Enabling it is what makes ServFlow serve that saved version, and enabling an already enabled agent again is how you apply later saved changes.
Related
Create an agent
Build an agent in the dashboard or from a config file, and call it from outside.
Entries and entry handlers
How a run starts, and where its reply goes.
Tools
Everything a sub-agent can call during a run.
Available actions
What steps and action tools run.