Agents
Found this helpful? Share it:
Found this helpful? Share it:
Agents are AI personas you design yourself. Give one a name, a personality, and instructions, then chat with it whenever you need its expertise. Each agent brings its own default model and its own memory 🤖
An agent is a named AI persona with custom instructions that shape how it behaves. Every conversation happens with a specific agent. That agent sets the personality, tone, and focus of the AI's responses.
You might create a "Code Reviewer" that focuses on finding bugs, a "Writing Coach" that helps with prose, or a "Research Assistant" that helps organize information.
Open a workspace and go to the Agents tab
Click New agent
Set a name and an icon or image, then write a system prompt that tells the agent how to behave
Pick a default model and an access preset, or keep the defaults and refine them later
Start simple. A short system prompt like "You are a helpful coding assistant who writes clean TypeScript" goes a long way, and you can refine it over time.
The system prompt is your main tool for shaping an agent's behavior. It is a set of instructions (up to 128 KB) that the AI follows in every conversation.
Tell the agent what to focus on, how to respond, what tone to use, and what to avoid. The more specific your instructions, the more consistent the agent's behavior.
The system prompt is re-sent with every message, so the agent keeps following your instructions even in long conversations.
Every agent carries a few settings of its own that you can change from its settings page:
Default model: the model a new chat with this agent starts on. You can still switch models inside any chat.
Its own memory: each agent you create gets its own private memory box (named after the agent) where it keeps notes across conversations. See Agent memory.
Default autonomy: how much the agent does on its own before checking with you. See Autonomy and confirmations.
When you start a chat without picking an agent, you talk to a built-in general assistant. It works well for everyday tasks: writing, research, coding, and general questions.
Once you build agents of your own, open a chat with the one that fits the task instead. Different agents can point at the same conversation topic with very different instructions.
Every agent belongs to a workspace. When you share a workspace with teammates, they get access to all the agents inside it.
Your whole team can then chat with the same agents, using the same instructions. Changes you make to an agent are visible to everyone with access.
There are a few ways to start a conversation with an agent:
Click an agent's name in the workspace Agents tab
Click an agent in the sidebar under Recent agents
Open its detail page and start a chat from there
Delete an agent from its settings or the Agents tab. It moves to the trash, where you can restore it for 30 days before it is purged for good.
Skills shape how an agent behaves in specific situations, like following a coding style guide or using a particular framework. Every agent automatically knows the skills in its workspace and reaches for the right one on its own, so there is nothing to attach.
Combine a focused system prompt with relevant skills for the best results, for example a "Documentation Writer" agent with a style guide skill.
Related articles
Agent permissions and tools
Control what your agents can reach and which tools they can use.
Autonomy and confirmations
Decide how much an agent can do on its own, approve or deny individual actions, and stop a run at any time.
Agent memory
How each agent keeps its own private memory across conversations, and how to read it, turn it off, or clear it.
Was this helpful?