Agent Analytics is Amplitude's new product for teams shipping AI agents into production, built to close the gap between offline evaluations and what actually happens once real users start asking questions. It captures agent traces, prompts, tool calls, responses, latency, and cost, then decomposes each turn into events that sit alongside standard product signals like session starts, page views, and button clicks. The launch matters now because AI coding assistants have made it trivial to ship agent features, but engagement dashboards flatten out the difference between a user who got a correct answer and one who got a confident hallucination.
The workflow moves from observing traces to scoring quality and failure modes, then tying those quality signals to conversion, retention, and revenue to see whether the agent actually moved the user forward, and finally running experiments and deploying personalized guides. Unlike observability tools that stop at the trace, Agent Analytics decomposes conversations into events that are directly queryable in the same funnels, cohorts, and retention analyses product teams already use in Amplitude , which lets an engineer ask questions like whether a model swap changed sign-up rates or whether hallucinated answers correlate with drop-off.
The release is part of a broader push into agentic analytics from Amplitude (Nasdaq: AMPL), announced by co-founder and CEO Spenser Skates alongside a Global Agent, four specialized agents for dashboards, session replay, web experimentation, and feedback, and MCP updates that pipe behavioral data into tools like Cursor, Claude, ChatGPT, Figma, Lovable, Notion, and GitHub. For founders and operators putting agents in front of customers, it is aimed at the specific blind spot between an eval suite passing and a support inbox filling up.
Comments (9)
Sign in to join the discussion.
P
Priya VatsalSep 3, 2026
The 'AI fails silently' framing is the whole pitch in four words. Whoever wrote that tweet deserves the rest of the week off.
K
Kwame Adu-BoahenSep 3, 2026
Every agent team I know is currently duct-taping Langfuse traces to a Notion doc. If this actually replaces that stack I owe someone a beer.
M
moonveilSep 3, 2026
Cool but I still haven't figured out regular analytics. Call me when there's a Google Analytics tab that just says 'agent good' or 'agent bad'.
E
Elena P.Sep 3, 2026
Curious how you're defining a 'session' when a single agent turn can span 12 tool calls and a human takeover. That taxonomy is where every vendor I've tried quietly falls apart.
T
Tomas LindqvistSep 3, 2026
The video cuts from dashboard to dashboard so fast I got motion sick. Would've killed for one slow scroll through an actual trace view.
B
Ben MarchandSep 3, 2026
Reply-guy energy but: the tweet buries the lede. 'Every agent looks engaged until you check the answers' should have been the hook, not the third line.
R
Rajshree K.Sep 3, 2026
Genuinely asking: what does retention look like for teams that adopt this in month one vs month six? Agent tooling churn has been brutal across the category.
H
Hajime OharaSep 3, 2026
Are the SDKs open source or is this a closed instrument-and-pray situation? Docs page 404s for me on mobile fyi.
Y
Yara LatifSep 3, 2026
Spenser, any chance you'd hop on a quick call this week? Working on a piece about the observability layer for agents and would love your read on where this market goes in 2025.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Amplitude
We help companies unlock the power of their products.
We help companies unlock the power of their products.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Agent Analytics is Amplitude's new product for teams shipping AI agents into production, built to close the gap between offline evaluations and what actually happens once real users start asking questions. It captures agent traces, prompts, tool calls, responses, latency, and cost, then decomposes each turn into events that sit alongside standard product signals like session starts, page views, and button clicks. The launch matters now because AI coding assistants have made it trivial to ship agent features, but engagement dashboards flatten out the difference between a user who got a correct answer and one who got a confident hallucination. The workflow moves from observing traces to scoring quality and failure modes, then tying those quality signals to conversion, retention, and revenue to see whether the agent actually moved the user forward, and finally running experiments and deploying personalized guides. Unlike observability tools that stop at the trace, Agent Analytics decomposes conversations into events that are directly queryable in the same funnels, cohorts, and retention analyses product teams already use in Amplitude , which lets an engineer ask questions like whether a model swap changed sign-up rates or whether hallucinated answers correlate with drop-off. The release is part of a broader push into agentic analytics from Amplitude (Nasdaq: AMPL), announced by co-founder and CEO Spenser Skates alongside a Global Agent, four specialized agents for dashboards, session replay, web experimentation, and feedback, and MCP updates that pipe behavioral data into tools like Cursor, Claude, ChatGPT, Figma, Lovable, Notion, and GitHub. For founders and operators putting agents in front of customers, it is aimed at the specific blind spot between an eval suite passing and a support inbox filling up.
Comments (9)
The 'AI fails silently' framing is the whole pitch in four words. Whoever wrote that tweet deserves the rest of the week off.
Every agent team I know is currently duct-taping Langfuse traces to a Notion doc. If this actually replaces that stack I owe someone a beer.
Cool but I still haven't figured out regular analytics. Call me when there's a Google Analytics tab that just says 'agent good' or 'agent bad'.
Curious how you're defining a 'session' when a single agent turn can span 12 tool calls and a human takeover. That taxonomy is where every vendor I've tried quietly falls apart.
The video cuts from dashboard to dashboard so fast I got motion sick. Would've killed for one slow scroll through an actual trace view.
Reply-guy energy but: the tweet buries the lede. 'Every agent looks engaged until you check the answers' should have been the hook, not the third line.
Genuinely asking: what does retention look like for teams that adopt this in month one vs month six? Agent tooling churn has been brutal across the category.
Are the SDKs open source or is this a closed instrument-and-pray situation? Docs page 404s for me on mobile fyi.
Spenser, any chance you'd hop on a quick call this week? Working on a piece about the observability layer for agents and would love your read on where this market goes in 2025.