Custom agents are the core of Gumloop; a good setup routine makes them reliable from day one.
A custom agent is one you build for a specific job, such as "Sales research" or "Support triage". Its configuration page (the builder) has four tabs: Agent, Triggers, Access and Settings.
When you create one, guided setup asks what the agent should handle and configures a first draft from your answer. Then follow Gumloop's recommended routine:
- 1. Add connectors for the apps the job needs, and nothing more.
- 2. Run a real task from start to finish in chat, correcting it as you go.
- 3. Ask the agent to write its own instructions from what worked, or turn the approach into a skill.
The instructions (often called the system prompt) are standing rules the agent reads before every task: its role, the steps it should follow, the tone, what it must never do, and when to ask a human. Be specific, as you would for a new colleague: "Reply in English, under 120 words, and never promise refunds".
Other sections of the Agent tab:
- Model: which AI model it uses (next lesson).
- Skills, Knowledge Sources, Subagents and Secrets: covered later.
- Abilities: built-in powers such as web search, web page reading, image generation, searching past conversations, asking you questions, and discovering tools. Turn off what the job doesn't need.
Changes only apply after you press Save (Ctrl/Cmd + S).
Write a clear name and description: other people, and other agents, use them to decide when this agent is the right one.
Remember
- Builder tabs: Agent, Triggers, Access, Settings
- Routine: add connectors, run a real task, then capture instructions or a skill
- Specific instructions and minimal abilities make agents reliable; remember to Save
Watch out for
- Writing long, vague instructions before ever running a real task with the agent.