Open Agents in the navigation rail and click New agent to launch a guided five-step builder. Each step takes you from a blank persona to a working agent — and you can test it live in the Try panel before sharing.
Steps build on each other, but you can jump back to any earlier step to refine it. Only an identity and a model are strictly required — the rest add capability.
Give the agent a name, a short description, and an icon so teammates instantly recognise what it does. This is how the agent appears in the agents list, in chat, and in the #-mention picker.
Write the instructions that define the agent's role, tone, and rules — its persona. You can compose saved prompts from the Prompt Library directly into this system prompt to reuse house style and guardrails.
Add a few example prompts. They appear as one-click suggestions when someone opens a chat with the agent, so new users know what to ask without a blank page.
The heart of the agent: choose the model and sharing scope; attach knowledge folders for retrieval; tune temperature, top-p, and max output tokens; attach skills; and grant the integration actions the agent may call.
Optionally add sub-agents the agent can delegate to as a supervisor, and define form fields to collect structured input when the agent is run as a fillable form or task.
The identity step is what makes an agent feel like a teammate rather than a setting. Pick a clear name (for example, Support Triage or Release Notes Writer), a one-line description of what it is for, and an icon. These show up everywhere the agent is referenced.
The system prompt is the agent's standing instructions — its role, voice, the rules it must follow, and the format you want answers in. Treat it like a job description. Because you can compose saved prompts from the Prompt Library into it, shared house style and safety language stay consistent across every agent your team builds.
A focused agent outperforms a vague one. State what the agent should do, what it should refuse, and how to format answers. If the agent should always cite its sources or always return a structured object, say so here.
Conversation starters are example prompts that surface as clickable suggestions when someone opens a chat with the agent. Good starters double as documentation — they show colleagues exactly what the agent is good at, so the first interaction is productive instead of a guess.
This step turns a persona into a capable agent. Here you make the decisions that govern how the agent thinks, what it knows, and what it can do:
| Setting | What it controls |
|---|---|
| Model | The LLM the agent runs on, chosen from your organization's six configured providers — OpenAI, Anthropic, Google, Groq, Ollama, and Ollama Cloud. The picker is data-driven, with current models such as claude-sonnet-4-5, gpt-4o-mini, gemini-2.0-flash, llama3.2, and gpt-oss:120b. |
| Sharing | Who can use or edit the agent — Private, Specific people & Agents, Workspaces, or Organisation, each with view or edit. |
| Knowledge folders | The folders the agent retrieves from. Attached folders feed the agent's retrieval so answers are grounded and cited. Knowledge & output. |
| Temperature, top-p, max output tokens | Sampling controls. Lower temperature and top-p make answers more focused and deterministic; max output tokens caps the length of each response. |
| Skills | Reusable instruction snippets that layer extra behaviour onto the agent without changing its model or tools. Skills. |
| Integration actions | The concrete tools the agent may call — send a Gmail draft, create a Jira issue, add a calendar event. Tools & actions. |
The final step adds two advanced capabilities, both optional:
The Try panel lets you chat with the agent live inside the builder, so you can check its tone, tools, and knowledge before you save or share. Every save is tracked in version history, so you can roll back if a change regresses behaviour — covered in Using agents.