Agents

An agent is a reusable AI assistant with a fixed identity: a name, a chosen model, a persona defined by a system prompt, and a curated set of tools. Instead of configuring each chat session from scratch, you build an agent once and anyone on your team can start a conversation with it instantly.

Agents can plan and execute multi-step tasks, calling tools in sequence, reasoning over the results, and continuing until the goal is reached — all within a single conversation turn.

Creating an agent

Open Agents in the left sidebar and click New agent. The agent builder opens as a form.

Name and description

Give your agent a clear name that tells team members what it does — for example, "Research Assistant" or "Support Bot". The optional description appears below the name in the agent picker and helps people choose the right agent for the job.

Choosing a model

Use the Provider and Model dropdowns to set which AI the agent always uses. Only providers you have connected in Providers appear in the list. Users chatting with this agent cannot switch models mid-conversation — the selection is locked to whatever you choose here.

See Models for guidance on matching models to task types. Fast, economical models suit high-volume, straightforward tasks; the most capable models are better reserved for complex reasoning.

Writing a system prompt

The system prompt is the agent's persona — an instruction block that is invisibly prepended to every conversation. This is where you define:

• The agent's role and area of expertise
• The tasks it should and should not perform
• The preferred tone and response format
• Any domain-specific rules or constraints

Be specific. A system prompt that says "You are a helpful assistant" produces generic behaviour. A prompt that says "You are a customer support specialist for Acme SaaS. You help users troubleshoot login and billing issues. Always ask for the user's account email before looking anything up." produces a consistent, focused agent.

Selecting tools

Under Enabled tools, check each tool the agent is allowed to use. Only tools you enable here are available during agent conversations — the model will not call tools that are not on the list, even if they are enabled globally. Start with the minimum set the agent genuinely needs.

All built-in tools and any dynamic tools you have created appear in this list.

Saving the agent

Click Save agent. The agent is immediately available to every member of your workspace — no extra sharing step is required. It appears in the agent picker at the top of the new-chat screen and in the sidebar under Agents.

Chatting with an agent

Start a new conversation and choose your agent from the picker, or click the agent's name directly in the sidebar. The conversation opens with the agent's model pre-selected and its tools already active.

As you chat, the agent plans autonomously. For a complex request it may call several tools in sequence — for example, searching the web, reading a linked document, and running a calculation — before delivering a final answer. Each tool call appears as a collapsible trace in the chat so you can follow the agent's reasoning.

Managing saved agents

All agents you have created appear under Agents in the sidebar. Click the overflow menu next to any agent to edit, duplicate, or delete it.

Workspace administrators can mark frequently used agents as pinned, which keeps them at the top of the agent picker for all team members.

Before changing a production agent's system prompt or tool list, consider duplicating it first. Test the duplicate in a private conversation and swap it in only once you are satisfied with the behaviour — this prevents disrupting ongoing team use.

Using agents in workflows

Agents can be used as steps inside a workflow — passing their output forward to subsequent steps or branching based on what they return. This lets you chain several specialised agents together for tasks that are too complex for a single agent to handle end-to-end.

Best practices

• Write precise system prompts — specificity is the single biggest lever for consistent agent behaviour. Iterate on the prompt whenever the agent produces off-target responses.
• Enable only the tools the agent needs — unnecessary tools increase the chance of unintended side effects and add noise to the model's decision about which tool to call.
• Match model to task — use a fast model for high-volume, simple tasks and reserve the most capable models for reasoning-heavy work to manage costs.
• Combine agents with workflows — for multi-step pipelines, orchestrate several focused agents in a workflow rather than overloading a single agent. See the Workflows guide.