Amplitude

Six months ago, we decided to treat our software factory like a system to optimize, not just a process to manage. We mapped the bottlenecks: 30 minute CI, 5 hour PR cycle time, manual review for everything. Then we fixed them one by one. Results: 3x PRs, cycle time down 7x, bug reports down 55% even as we shipped more. And 5% of our PRs now come from non-engineers. The biggest surprise: Fixing the foundations for agents made things better for engineers too. What’s good for agents is good for humans. Full writeup here, and bonus! We also turned it into a game. https://amplitude.com/3x

So what's the deal with this Amplitude / Statsig partnership? We're building safe rollouts and measurement for the AI age. Whether you're building probabilistic products or managing the acceleration of commits from coding agents, we are building the system to safely learn and grow from every change to accelerate impact.

Most teams ship agent personalities by accident. We designed Amplitude Global Agent’s on purpose. Jacob Newman shares his design decisions, starting with the two traits that had the biggest impact on whether users kept coming back: (1) Inquisitiveness: We initially told Global Agent not to ask follow-up questions, assuming that would make users bail. Turns out, we were wrong. An A/B test against a more inquisitive agent showed that users had longer conversations and were more likely to save the resulting analysis, so we decided to dial inquisitiveness up. (2) Helpfulness: We received internal feedback that the agent was giving instructions instead of actually doing the requested task, but we didn’t know how prevalent this issue was. So we built an Agent Analytics evaluator that read every conversation and flagged the pattern. It showed up in ~3% of all conversations, which was enough to justify dialing helpfulness up. Any team building an agent can run a similar loop to fine-tune its personality, instead of letting it become a byproduct of the underlying model or system prompt. Read more about Global Agent’s personality design → https://lnkd.in/g-h_VpAh

How do you run a feature launch when the data, flag controls, and team decisions all live in different places? Slowly, that's how. Tommy Keeley shares a workflow that puts all three into a single Slack channel using Amplitude AI Agents and Feature Experimentation. What your team can do from inside the launch channel: - Ask data questions in plain language, and Global Agent answers in the thread. - Ramp the feature flag with a single message instead of DM'ing eng. - See session replays of what users experience (rage clicks, broken flows, unexpected errors, etc.) when metrics drop. - Get the post-launch readout, including qualitative themes from surveys, support tickets, and sales calls. Read the full workflow → https://lnkd.in/gDmzqz-H

For my AI week project, I pointed a Ralph loop at Claude Code, gave it my backcountry ski planning app, and walked away. By Friday it had shipped 102 features. Slope-angle overlays on a 3D Mount Hood. A physics-based avalanche runout simulation. A mushroom foraging prediction model I never asked for. The loop itself wasn't the interesting part. What made it work was the feedback cycle: each feature the agent built instrumented itself with Amplitude events and session replay. The next cycle had real behavioral signals to work from — what got used, what broke, where sessions got stuck. Cycle 10 was meaningfully smarter than cycle 1. Three things I'm carrying forward: The dispatcher matters more than the loop. Any while-loop will spin. Only one with honest outcome signals will converge toward value. Self-instrumentation is what makes it compound. Without telemetry, every cycle runs on the same priors as the first. The bottleneck moves. When execution gets cheap, taste and prioritization become scarce resources. The loop did the building. I did the judging. I wrote up the full article including the parts that almost didn't work 👇 https://lnkd.in/gGK7B6kF Big Shoutout to Amplitude co-conspirators Frank Lee William Newton Peyton Rose Adam Struck Richard F.

We’ve been prototyping a “self-improving product” agent on Amplitude and during AI week, Eric Carlson gave the agent a product idea, hooked it up to Amplitude, then the product autonomously built itself out based on feedback from how users interacted with the product 🤯 As coding agent capabilities continue improving, the dream of self-improving products is already becoming a near-term reality. However coding agents require high-quality signals on what’s working and not working, context and goals that the agents should continue optimizing for, and task clarity on the features that are valuable and worth working on (or saying no to!). We have a ton more to share in the coming weeks about this agent orchestration system and I'm excited to bring Amplitude directly into agentic coding workflows to enable self-improving software!

We are starting to open a number of roles at Amplitude in Data Science and Engineering with the new Statsig partnership. Please reach out if you have a background in Experimentation. More to come.

After a strong Q1 and an intense AI Week at Amplitude , one thing is clear : the best leaders don’t just talk about balance, they live it. Taking well‑deserved time off with family and friends isn’t a pause on performance. It’s an investment in clarity, energy, and long‑term impact. When we honor our personal lives, we come back to our teams more grounded, more focused, and more human. As leaders, our job is to lead by example and embody the values we want to see in our culture, rest, reflection, and respect for time with loved ones included. What’s one thing you do to recharge after an intense period?