Custom Agents

Automate the recurring analysis you'd otherwise repeat by hand each week by building your own AI agent. You describe what the agent does, give it access to your Amplitude data and connected tools, and put it on a schedule, so it monitors releases, grooms a backlog, tracks a funnel, or summarizes customer research on its own. Product managers, analysts, and technical teams who run the same workflow again and again use Custom Agents to do that job for them.

Build a custom agent when no Specialized Agent already covers the workflow you want to repeat. For one-off questions and ad-hoc exploration, use the Global Agent instead.

How Custom Agents work

A custom agent is a saved, named configuration: a set of instructions, a model, the tools it can use, and an optional schedule and delivery destination. When the agent runs, it follows your instructions against live Amplitude data plus any external tools you connect, then returns its findings in Amplitude and, if you configure it, in Slack or email.

Custom Agents sit alongside the other agents in Amplitude AI and complement them:

Global Agent is the always-on chat entry point for ad-hoc exploration and one-off questions.
Specialized Agents are the agents Amplitude builds and tunes for a specific job, such as the Dashboard Agent, Session Replay Agent, Customer Feedback Agent, and Website Conversion Agent.
Custom Agents are the agents you author yourself when no built-in agent matches the workflow you want to automate.

Use cases

Teams across product, engineering, and go-to-market build custom agents to automate recurring work:

Product management: Cluster customer feedback into themes and file ranked Jira or Linear tickets, or detect experiments that reach significance and draft a decision doc.
Engineering: Run a post-deploy release health check that compares each feature's funnel, error rate, and adoption against baseline and files a regression ticket to the author, or find feature flags stuck at full rollout and file cleanup tickets to the owning team.
Customer success and marketing: Flag strategic accounts with slipping engagement and alert the owner with the drop-off point, or detect newly dormant users, build the cohort, and draft a win-back campaign for approval.

Prerequisites

Amplitude AI enabled for your organization. Amplitude enables it by default unless your organization opted out of AI features.
Access to the data the agent analyzes. An agent inherits your Data Access Controls and permissions, so it can only use data you can already access.
For external tools, an authorized connector for each app you want the agent to use.

How to create a custom agent

You can create a custom agent in three ways:

From the builder: In Amplitude, go to Agents, then create a new agent and configure it from scratch.
From chat: Turn a Global Agent conversation into a saved agent, or type the /create-custom-agent command followed by a description to generate a draft.
From a template: Start from a prebuilt template in the Agent Library, such as churn monitoring or feature adoption tracking, and adjust it for your workflow.

Each path produces a draft you can edit before you publish it.

How to configure a custom agent

Open the agent in the builder to set up how it behaves:

Name and description: Identify the agent and describe what it does.
Instructions: Write natural-language instructions that tell the agent its goal and how to approach it. Reference Amplitude assets with an @mention to pull a specific chart, dashboard, or cohort into the agent's context.
Model: Choose the model that powers the agent.
Tools and connectors: Give the agent access to Amplitude tools and to external apps you connect through Agent Connectors, such as Slack, Jira, Linear, GitHub, Notion, Confluence, Granola, and Sentry. Connectors let an agent both read from and write to those tools, so it can investigate across your stack and then act, like filing a ticket or updating a doc. The builder shows only the connectors your organization authorizes, and an agent reaches only the data you can already access in each connected tool. If a connector you need isn't available, you can create a custom connector.
Capabilities: Turn on the capabilities the agent needs:
- Memory: Carry context across runs.
- Customer Context: Use product and business context you provide.
- Long Conversations: Support extended, multi-step runs.
- Extended Thinking: Allow deeper reasoning on complex tasks.

The more specific your instructions and context, the more accurate and useful the agent's output. Add two or three detailed lines about the product area, goals, and metrics that matter.

Connectors and the Amplitude MCP server work in opposite directions: connectors let an agent query your other tools from inside Amplitude, while the MCP server lets other tools query Amplitude data. Amplitude doesn't store the data an agent reads through a connector.

How to schedule a custom agent

Put a published agent on a recurring cadence so it runs unattended. Set the agent to run daily or weekly, and choose the day, hour, and time zone. At each run, the agent analyzes your data and pushes its summary and recommendations to the destinations you configure.

Deliver results to Slack and email under the agent's notification settings.

To deliver findings to Slack, invite the Amplitude Slackbot to the target channel. Without it, the agent can't post to the channel.

Amplitude creates custom agents on a per-user basis, and each agent respects your individual permissions. By default, custom agents are private to the user.

A draft agent is in progress while you build and test it.
You can share a published agent with your organization, so teammates can view its insights and reports.

Manage your agents from the Agents page, where the All Agents tab lists every agent you can access and the My Agents tab lists the ones you created. As the owner, you control the agent's schedule and configuration.

For agents that recommend actions, a human reviews and approves any change that affects end users, such as launching an experiment or creating a cohort, before Amplitude runs it. Admins can view and audit agent input and output across the organization.

Was this helpful?