A few weeks ago, one of our designers, Will Newton, built a Custom Agent to take on a job that never quite fit into the week. Every Monday morning, it reviews the prior week's session replays, finds the top friction points where users got stuck, validates each one against event data to size how many people it hit, and writes up the highest-impact ones with the evidence behind them. Then it files the Linear tickets and hands them off to start the work. The first run came back with five tickets scoped to real problems.
The part worth paying attention to isn't that an AI wrote some tickets. It's what the work was grounded in. The agent didn't reason about which users were probably affected. It read what they did, in the product, during the window the bug was live, and built from there.
That distinction is the whole point.
Most AI agents are good at helping you think. You bring a question, they help you reason toward an answer. That's useful, and it's also where most of the market has settled: a faster way to hypothesize. The work that actually moves a product forward starts somewhere else. It starts with what your users did, and it ends with something happening because of it.
That's the gap Custom Agents is built to close.
A Custom Agent is a configured worker you set up once inside Amplitude. You give it a name, write instructions describing the job, choose the model that fits the task, and connect it to the tools it needs to read from and act in. Then it runs the same job, the same way, on a schedule or on demand.
If you already use Amplitude's Global Agent, you're partway there. Global Agent is the chat interface where you ask a question and get an answer, and it's also where you build your first Custom Agent. The shift is from asking to standing work. Instead of bringing a question each time, you set up an agent once and it runs the same job on its own. You find the output waiting instead of remembering to go get it.
What makes the output trustworthy is where it comes from. Every Custom Agent runs on real behavioral data from Amplitude, your actual charts, cohorts, and events, plus context from the systems your team already uses through Agent Connectors. It connects to Slack, Jira, Linear, and more, so an agent can read a PRD, pull the matching usage, and file the follow-up ticket in one pass. The result reflects what happened across your product and your stack, not a confident guess assembled from whatever was in the prompt.
Will's P0 triage is one shape of this. Most of them have nothing to do with engineering.
Every Monday, someone assembles the leadership digest. WAU from Amplitude, product wins from Linear, incidents from Jira, formatted and posted before standup. The night before a customer call, someone pulls a pre-call brief: usage trends, adoption changes, open tickets, stitched together from three tools at midnight. After a campaign, someone reconciles spend against activation. When a flag rolls out, someone should compare the treated and control cohorts and report back, and usually does, eventually.
None of this takes expertise. It needs someone to remember, find the tools, pull the data, and put it where it belongs.
The signal already exists. Getting it to the right person, in the right format, before the moment passes is the part that keeps falling on a human. A Custom Agent takes that standing job and runs it. And because it can act in connected tools, it doesn't stop at the analysis. It builds the cohort, drafts the outreach, files the ticket, updates the doc. The loop between knowing and doing closes without a handoff.
Custom Agents launches with a library of agents our own teams already run, so you're not staring at a blank prompt. The Specialized Agents you may know from earlier this year carry forward here as templates rather than going away, and the library grows over time, eventually with agents customers build and share.
Custom Agents is open beta and available to all customers starting today. Explore the Custom Agent Library.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。