AI Agent¶
One agent: a model, a system prompt, and the tools it may reach. Everything an agent is made of is a row. The prompt is a column, the tool list is a table, and the specialists it may delegate to are another table. Nothing about an agent is a constant in Python code.
This app builds an agent from these rows. It does not run one. A consuming app runs it.
Fields¶
| Field | Type | Required | Description |
|---|---|---|---|
name |
String | Yes | Unique. Also the name a supervisor calls this agent by, unless a binding overrides it. |
description |
String | No | What this agent is for. Read the caution below before you write it. |
system_prompt |
Text | Yes | The standing instructions of the agent. |
model |
AI Model | Yes | The chat model. An embedding model is refused. |
pattern |
Choice | Yes | single, subagents, or skills. |
enabled |
Boolean | Yes | A disabled agent is not built. |
temperature |
Decimal | No | Overrides the model. |
num_predict |
Integer | No | Overrides the model. -1 means unlimited. |
max_iterations |
Integer | Yes | How many model calls one run may spend. |
tenant |
Tenant | No |
The description decides whether the agent is called¶
CAUTION: A supervisor reads the description of a specialist to decide whether to delegate to it.
A description that reads badly to the model produces no error, no exception, and no log line. It
produces silence. When a multi-agent system returns an empty answer, examine the name and the
description first.
Write the description the way you write it for a new colleague on the first day. Say what the agent does. Say what to send it. Keep it to one or two sentences on one line. A bulleted, multi-line description stopped a tool being called at all in every measured run.
Three settings resolve in one chain¶
temperature and num_predict are read from the agent first, then the model, then the provider.
The first one that carries a value wins. This lets one model serve a supervisor at temperature 0
and a writer at temperature 0.7.
Choose a pattern¶
Start with single, and give the agent every tool. Measure it. Most systems stop there.
Move to subagents when one prompt can no longer hold every rule. A specialist gets an isolated
context and its own prompt. You pay for that with more model calls, every run.
Move to skills when the rules are large and only advisory. See AI Skill for the
measured weakness of that pattern.
Related¶
- AI Agent Tool binds a tool to an agent.
- AI Agent Subagent binds a specialist to a supervisor.
- AI Agent Skill binds a skill to an agent.
- AI Agent Thread records one run.