Build an agent in the UI

The Build tab creates an agent inside LangSmith instead of in code. You fill in a configuration form, and LangSmith creates a no-code agent from it when you submit. Nothing is created until you do. Use it when you want a working agent without a local project or a deployment. Everything it produces is an agent like any other, so its traces, dashboards, and evaluators behave the same as one you instrumented yourself. The Build tab appears only in an agent-based workspace. To build an agent from files on your own machine instead, see Managed Deep Agents.

​

Create the agent

To create an agent:

1. In the left navigation, click Build. With no agent selected, Build asks you to pick one and offers New agent below the list. That list holds only agents built here, so an agent you traced from code does not appear in it.
2. Fill in the configuration sections. A line under the form’s title reads Nothing is created until you click Create Agent.
3. Click Create Agent.

LangSmith creates the agent, then finishes the rest of the setup, such as connecting accounts and registering schedules. A dialog reports what landed and offers Go to Studio to test the agent. Cancel leaves without creating anything and returns to the list of agents. If part of the setup does not finish, the form reports which part and offers Retry unfinished changes, so a second attempt does not create a second agent. + Agent above the agent list also reaches this form. It opens the New Agent panel, whose Builder option, “Build and configure inside LangSmith”, opens this tab on a blank form. For the panel’s other three options, see Create an agent.

​

Configure the agent

The form is one page of collapsible sections. Every section starts open, and the browser remembers which sections you collapse.

• Basics: The agent’s Name, a Description of what the agent does, and the Model that powers its responses. The name is required, and it is the label you see in the agent switcher. The model is required when you create the agent.
• Instructions: Agent instructions, covering the agent’s role, its responsibilities, and how it should work. LangSmith saves this exactly as entered.
• Tools and accounts: The connections the agent uses for its tools, added with Add connection. On a new agent, the accounts are granted when you create it, and adding a workspace connection starts its authorization flow right away. For whose credentials the agent uses, see Agent identity.
• Skills and memory: The skills the agent pulls in for specific tasks, and the memory files it keeps between runs. Update memory and instructions lets the agent revise its own configuration, either applied automatically or held for your approval.
• Channels: Whether the agent works in Slack. On a new agent, you authorize Slack after you create the agent. On an existing agent, connecting or removing Slack takes effect right away. Where Slack setup is unavailable in the workspace, the section says so.
• Schedules: The scheduled runs the agent makes on its own.

Once the agent exists, the state line under the title reads Unsaved changes or All changes saved.

​

Change the configuration later

Select the agent in Build to reopen its configuration, now headed Agent configuration. Edit any section and click Save changes. Without edit access on the agent, the form opens View only. Renaming needs permission to update the agent’s tracing projects, and the Name field says so when you lack it.

​

Find the traces

An agent built here is created with a Production environment alone, and its runs land there. It gains other environments only when something reports into them. To see its environments, select the agent and open Tracing, or read the summary on the agent Overview. An agent built in the UI has no deployment, because it runs on a Studio runtime instead of one you deployed. It has no Deployments section on its Overview, and no deployment can be created for it, so the agent environment deploy flags do not apply to it. The workspace’s other deployments are hidden while that agent is selected.

​

See also

• Agents
• Navigate agents
• No-code agents
• Agent environments

---