LangSmith Agent Builder Tutorial (No-Code): Build an Email Triage Assistant

LangSmith’s new Agent Builder (private preview) is a no-code way to turn a natural language idea into a production-ready agent. Instead of writing workflows, you describe what the agent should do, and Agent Builder drafts the system prompt, proposes tools, adds triggers, and supports human-in-the-loop interrupts, persistent memory, and self-updates. 

You can compose sub-agents for specialized tasks, connect Google integrations (Docs/Sheets) alongside web search, and then observe everything with built-in Tracing and Monitoring dashboards.

In this tutorial, we’ll build an Email Triage Assistant end-to-end. You’ll learn how to:

• Access the Builder and create a new agent from a natural-language brief
• Review the generated plan and approve actions
• Authenticate providers and deploy the agent
• Test real emails, and iterate on instructions
• Tracing and monitoring help us to inspect runs, latency, and token costs

By the end, you’ll have a working agent that classifies incoming email, drafts replies, and logs daily actions to Google Sheets, without writing any code.

If you’re keen to learn more about building AI agents, I recommend checking out the Multi-Agent Systems with LangGraph course. 

What is LangSmith Agent Builder?

LangSmith’s Agent Builder (preview) gives you a guided way to design agents with memory, sub-agents, triggers, and human review, all from a single UI. Under the hood, Agent Builder uses deep-agents, an open-source agent harness built on LangGraph that compiles your natural language brief into a stateful agent rather than a hand-drawn workflow. 

It plans before executing (to manage long tasks and cost), delegates to sub-agents for specialized work, taps a tool layer via middleware like (files, todos, optional shell, web/MCP), supports human-in-the-loop interrupts for sensitive actions, and runs as a compiled LangGraph graph for streaming, tracing, and studio integration.

Key capabilities of this Agent Builder are:

• Natural language setup: Builder drafts instructions and wiring based on just natural language input.
• Triggers: These start runs from external events (for example, new email received!) or schedules.
• Tools: One-click auth for Gmail, Google Docs/Sheets, Calendar, Web Search, and more.
• Sub-agents: It breaks up complex tasks into focused helpers like research, summarization, search, and more.
• Human-in-the-loop: This loop enables humans(users) to interrupt the builder so actions pause for your approval.
• Memory and self-updates: Persistent memory and controlled self-edits help to update instructions and tools.

The result is a no-code experience that still gives you observability, approvals, and iteration hooks. To get access to LangSmith  Agent Builder, you need to sign up for the waitlist. 

LangSmith Agent Builder Example: Build an Email Triage Assistant

We’ll build a simple assistant for a freelancer who receives many agency emails. The agent will:

1. Read new emails and classify them into reply now, schedule, delegate, or archive.
2. Draft a 1-sentence summary along with a 2-line reply suggestion.
3. If the sender/domain is unfamiliar, then it will add a 1-line context via web search.
4. Log everything to a Google Sheet titled “Today — Email Actions (YYYY-MM-DD)”.
5. Get user approval before creating or updating documents.

Step 1: Access Agent Builder

Once you’re off the waitlist, sign in to LangSmith and open Agent Builder (Beta) from the left sidebar.

You’ll land on the No-code Agent Builder page. Click + New Agent to start a new agent, or choose Learn More to read the docs before you begin.

Step 2: Describe your agent

Start by typing a concise brief and choosing a model. You can build an agent manually as well, but that’s a topic for another blog. 

Choose from the listed models or add a custom one to fit your use case. 

Here is the prompt that I used:

You are an email triage assistant. When a new message arrives, classify it as {reply now, schedule, delegate, archive}. Draft a 1-sentence summary and a 2-line reply suggestion. If the sender/domain is unfamiliar, add a 1-line context via web search. Update (or create) a Google Doc titled “Today — Email Actions (YYYY-MM-DD)” with columns: Sender | Subject | Action | Suggested Reply | Due. Pause for approval before adding or modifying the doc.

If you select Sonnet 4.5, you’ll be prompted to add your ANTHROPIC_API_KEY. Enter the key and save.

After you submit the description, the builder analyzes your intent, proposes tools/triggers, and drafts the initial instructions automatically.

Step 3: Follow-ups

After the builder inspects the available tools and triggers, it will ask a few targeted follow-up questions like preferred destinations for outputs, approval requirements, labeling rules, etc.

Your responses are used to finalize the agent’s instructions, attach the right tools/triggers, and continue the setup automatically.

Step 4: Review the generated plan

After you answer the follow-ups, the builder shows a proposed graph including triggers, agent, and any connected tools (like Gmail, Google Docs) already wired together.

Use this screen to sanity-check the design before creation:

• Trigger: When and how the agent starts (email received, schedule, etc.).
• Agent: Model choice and core instructions.
• Tools: What the agent can call and for which actions.

Think of this step as a pre-creation preview and a quick way to validate scope, data access, and guardrails. If it looks right, proceed. Next, you’ll authenticate and connect tools so the agent can actually act (send drafts, update Docs, etc.).

Step 5: Authenticating and connecting

Before your agent can read emails or write to a Google Sheet, you need to grant it access. 

In Agent Builder, this is a guided, one-time process. You’ll connect Gmail and the google-langsmith-prod integration (for Google Sheets actions). 

Once connected, you can safely require human approval before any write action.

Step 5.1: Connect to Gmail

To connect your agent to your Gmail account, follow these steps:

• In the Connections panel, click Connect next to Gmail — Email Received.

• Sign in with the Google account you want the agent to monitor and approve the permissions.
• Back in Agent Builder, click I’ve completed authentication.

Once the Email Received trigger is set, your agent can classify and respond to incoming messages according to your instructions.

Step 5.2: Authenticate to google-langsmith-prod

Next, authenticate google-langsmith-prod in a similar way as above. Here are a few steps that you can follow:

1. Click Connect next to google-langsmith-prod.
2. Approve access for Sheets with the same (or intended) Google account.
3. Confirm in Agent Builder.

Now, your agent can create/update Google Docs and Sheets (like append rows, update tables, or generate a new sheet, etc) while still honoring any interrupt/approval steps you configured.

With both connections active, click Create Agent. You’ll see the completed graph and can start testing end-to-end.

Your agent now has the minimum, secure set of permissions to receive emails and write action items to Google Sheets. Next, run a few test messages and use interrupts to approve or edit any draft updates before they’re applied.

Step 6: Testing the agent

Start with a few real emails to see how the agent classifies and responds.

Example 1:  Informational email 

I copied and pasted a routine “YouTube information” message in the test chat window, and the agent correctly tagged it as low priority and suggested an “archive/no reply” action with a one-line summary.

Example 2: Time-sensitive email 

For this example, I pasted an urgent note that required action. The agent classified it as reply now, produced a concise summary, and drafted a two-line response as well. 

It addressed the reply to the wrong recipient (minor routing error), but did identify the right subject and intent, which is quite useful for a quick edit and send. 

The best part about this tool is that it requires approval before performing any action on your behalf. This human-in-the-loop step let me fix the addressee and confirm updates, ensuring no unintended changes.

Step 7: Deploy the agent

So, we’ve tested the flow and reviewed traces; now it’s time to put the agent to work.

Deploying moves you from a controlled test to live inputs, while keeping human-in-the-loop approvals so nothing executes without your say.

Click Use Agent to switch the agent into active use.

Here is what I asked the agent.

Look through the top 10 emails and suggest action items for each in the form of a table.