Creating an Agent
How to configure an agent on top of an LLM provider and start chatting with it.
An agent is a configured, reusable instance of a model. It pairs a specific model from one of your LLM providers with a name, description, and instructions, so you can shape its behavior for a particular purpose without writing any code.
Providers vs. agents
A provider is just a connection to a backend. An agent is what actually gets used in a chat - it's the combination of a provider, a model, a name, and instructions.
Prerequisites
Before creating an agent, make sure you have at least one LLM provider already configured.
Creating an agent
Open the Agents section
Navigate to Settings → AI → Agents and select Add Agent.
Describe the agent
- Name - how this agent will appear when starting or switching chats (e.g.
Support Assistant,Code Reviewer). - Description - a short summary shown alongside the agent in chat pickers. Use this to help colleagues choose the right agent at a glance (e.g. "Answers questions from the internal HR handbook").
Write the instructions
The Instructions field is where most of the customization happens. It tells the model how to behave: tone, constraints, output format, domain knowledge - anything that should hold across every conversation.
Tips for effective instructions:
- State the agent's role explicitly ("You are a support assistant for...").
- List hard rules ("Never invent policy numbers", "Always answer in German").
- Describe the desired output format when it matters (bullet points, tables, JSON).
Select a provider and model
Choose one of your previously configured providers. The list of available models is then retrieved dynamically based on the provider type:
- Cloud providers (OpenAI, Anthropic, Gemini, DeepSeek) - the platform calls the provider's API to fetch the current list of models available to your account.
- Self-hosted providers (Ollama, llama.cpp) - the platform queries your server directly and lists whichever models are already pulled or downloaded on that instance.
No models showing up?
For self-hosted providers, an empty model list usually means no models have been downloaded on that server yet. Pull or download a model on the external instance first, then refresh the list.
Configure capability toggles
Three switches control what the agent can do:
- Tools - allow the agent to use tools (built-in platform tools and any tools exposed by your connectors). When enabled, you can restrict exactly which tools are available via the tool catalog selector that appears below the toggle.
- Vision - allow the agent to accept image attachments in chats.
- Reasoning - mark the model as a reasoning model where the underlying provider supports it.
Model compatibility
Not every model supports every capability. If a toggle doesn't seem to have an effect, verify that the selected model actually supports tool calling, vision input, or reasoning on your chosen provider.
Save the agent
Save the agent - it becomes immediately available for starting new chats.
Multiple agents, one model
You can create as many agents as you like on the same underlying model, each with different instructions. For example, two agents both running claude-sonnet-5 - one configured as a terse internal documentation assistant, another as a friendly customer-facing support bot - are entirely independent from each other and can be edited without affecting one another.
Next steps
Head over to Chatting with Agents to start your first conversation, or browse Tools to give your new agent hands.