--- title: "Agentic investigation with bitdrift: Why is my conversion rate dropping?" slug: "agentic-investigation-conversion" blurb: "Let your agents query reality. Use the bitdrift CLI with your agent of choice to answer questions about your mobile app with full-fidelity observability." metaDescription: "Let your agents query reality. Use the bitdrift CLI with your agent of choice to answer questions about your mobile app with full-fidelity observability." cover: url: "/assets/posts/agentic-investigation-conversion/feature-blog_hero_image_33-1-desktop@1x.webp" alt: "Agentic investigation with bitdrift CLI" socialThumbnail: url: "/assets/posts/agentic-investigation-conversion/feature-blog_hero_image_33-1-desktop@1x.webp" alt: "Agentic investigation with bitdrift CLI" author: - "collin" tags: - "mobile" - "android" publishedDate: "2026-06-19T12:00:00.000Z" modifiedDate: "2026-06-19T12:00:00.000Z" --- If you're a mobile engineer, the following scenario probably sounds familiar. It's a Monday morning. You open your analytics dashboard and notice that your purchase funnel conversion rate dipped over the weekend. So you spend the next two hours opening tabs, running queries, and cross-referencing logs trying to figure out what changed. The question sounds simple. The answer never is. The investigation loop is exactly the kind of thing an agent should be doing for you: 1. Authenticate 2. List apps 3. Find the right workflow 4. Pull the chart data 5. Check for errors 6. Compare time windows With the [bitdrift CLI](https://docs.bitdrift.io/cli/quickstart.html) and an agent like Claude Code, it can. ## Unpacking the agentic investigation, step by step Let's walk through exactly what happened in the demo so you can replicate it in your own environment. ### Step 1: Load the skill Inside Claude Code, the `/bd-cli` skill loads the bitdrift CLI tool definitions and gives the agent access to all platform commands. No bespoke prompt engineering required; the skill handles context about available commands, authentication patterns, and output formats. ``` > /bd-cli The bd CLI skill is loaded. What would you like to investigate or do with the bitdrift platform? > Why is my conversion rate dropping? ``` ### Step 2: Authenticate and discover The agent calls `bd auth --status` to verify credentials, then `bd app list` to enumerate available apps. It finds one app — `ai.bitdrift.shop` on Android — and immediately pivots to finding conversion-related workflows and charts. ``` $ bd auth --status -ojson 2>/dev/null { "auth_type": "jwt", "authenticated": true } $ bd app list -ojson 2>/dev/null { "platforms": [ ... ] } // One app found: ai.bitdrift.shop (Android, v1.0) ``` ### Step 3: Find the right workflow The agent lists all workflows and filters for anything funnel-related. It finds the Shopping Funnel workflow and extracts its ID and step definitions — Categories, Browse Categories, Cart, and Payment — before pulling the chart data. ``` $ bd workflow list -ojson 2>&1 | head -50 // "Shopping Funnel" — that's directly relevant { "id": "gYM", "name": "Shopping Funnel", "state": "LIVE" } $ bd workflow charts IHRK --app-id ai.bitdrift.shop \ --platform android --last 7d -ojson 2>/dev/null ``` ### Step 4: Read the funnel data With seven days of chart data in hand, the agent compares step-level performance. Here is what the funnel looked like at the time of investigation: | **Categories** 969 entry | **Browse** 969 100% | | --- | --- | | **Cart** 607 63% | **Payment** 588 97% | The significant drop, a 37% falloff, happens between Browse Categories and Cart. The agent flags this as the area to investigate further and issues commands to pull error and crash data for that screen over the same window. ### Step 5: Check for correlated errors ``` $ bd issue group list --app-id ai.bitdrift.shop \ --platform android --last 7d -ojson \ -jq '[.issue_groups] | {id, reason: .metadata.reason, users: .state.user_cou...}' $ bd workflow charts gGLc --app-id ai.bitdrift.shop \ --platform android --last 7d -ojson | python3 -c \ "import json, sys; ..." ``` By combining funnel chart data with issue groups and crash telemetry, the agent can surface whether the conversion drop correlates with a specific error condition. And it does it in a single, continuous investigation thread rather than across 5 separate tools. ## Why full-fidelity data changes what's possible Traditional observability [relies on sampling](https://blog.bitdrift.io/post/observability-3-0). Capture 1% or 10% of events, aggregate them server-side, and surface summaries in a dashboard. That approach works well enough for monitoring broad trends, but it falls apart for the kind of granular, per-session investigation that actually answers "why." When you ask an agent to investigate a conversion drop, it needs to query specific sessions, specific device conditions, and specific error states. Sampled data cannot reliably support that. You end up with a picture of what might have happened across a subset of users, not what actually happened on real devices. [bitdrift](http://bitdrift.io) captures events at full fidelity on device and streams them with negligible overhead. That means when the agent queries the Cart screen for error-correlated sessions over the last 7 days, it is querying every relevant session, not an approximation of them. **Agentic investigation only works when the data layer it queries is complete.** Sampled telemetry produces sampled answers. Full-fidelity observability [lets an agent query reality](https://blog.bitdrift.io/post/query-reality-ai-observability). ## Setting this up yourself To run an investigation like this in your own project, you need three things: 1. The bitdrift CLI installed and authenticated 2. The /bd-cli skill loaded in your agent of choice (I used Claude Code) 3. At least one workflow configured in the bitdrift platform for your app Workflow setup is the only step that requires upfront configuration. Once you have a funnel or exploration workflow in place, the agent can discover it automatically and query its charts without any additional instructions from you. The skill handles the rest — authentication, app discovery, command construction, and result interpretation. If you haven't set up the bitdrift CLI yet, the quickest path is to authenticate via `bd auth login` and then run `bd app list` to confirm your apps are visible. From there, the demo above is directly reproducible. ## Ready to run your own agentic investigation? [Get started with the bitdrift CLI →](https://docs.bitdrift.io/cli/quickstart) [Learn more about bitdrift.ai →](https://blog.bitdrift.io/post/query-reality-ai-observability) --- ## Video transcript Hey, I'm Collin from bitdrift. One of the most common questions we hear from mobile engineers is: why is my conversion rate dropping? It's a deceptively simple question that usually requires a lot of manual investigation across tools. Today I want to show you how the bitdrift CLI makes that investigation something an agent can do on your behalf. I'm going to launch Claude Code and load the bitdrift CLI skill with the /bd-cli command. Once that's loaded, I just type my question in plain English: Why is my conversion rate dropping? The agent starts by authenticating with the bitdrift platform, then it lists available apps and finds our Android app: ai.bitdrift.shop. It discovers a Shopping Funnel workflow that's directly relevant, so it pulls the chart data for that workflow and checks performance over the last seven days. Now, here's what makes bitdrift different. Because we capture events at full fidelity, not sampled, the agent is querying real device reality. It can describe every step of that funnel — Categories, Browse Categories, Cart, and Payment — and compare step conversion rates against the prior period to see where the drop-off happened. The agent also checks for errors and crashes that might explain the drop. It issues a command to list issue groups and charts the crash data alongside the funnel data so you can see both at once. At the end of this investigation you don't get a pile of raw data to interpret, you get an answer. The agent surfaces where conversion broke down, which step, which device condition, and whether there's a correlated error. That's the power of pairing an agentic workflow with full-fidelity mobile observability. --- ## Frequently asked questions ### What is agentic investigation in mobile observability? Agentic investigation is when an AI agent autonomously runs a multi-step diagnostic workflow on your behalf, authenticating, querying data sources, comparing time windows, and surfacing a conclusion, rather than requiring you to manually open dashboards and cross-reference tools. With the bitdrift CLI and an agent like Claude Code, you can ask a plain-English question like "why is my conversion rate dropping?" and receive a structured answer. ### How does the bitdrift CLI work with AI agents like Claude Code? The bitdrift CLI exposes platform commands that AI agents can call directly. When you load the /bd-cli skill inside Claude Code, the agent gains access to commands for authentication, app discovery, workflow listing, and chart retrieval. It can chain these commands together autonomously to investigate a specific question without additional prompting from you. ### Why does full-fidelity data matter for AI-powered investigation? Sampled telemetry, where only 1-10% of events are captured, produces approximate answers. When an AI agent investigates a conversion drop, it needs to query specific sessions, device conditions, and error states. bitdrift captures events at full fidelity on device, which means the agent is querying every relevant session, not a statistical subset. Sampled data produces sampled answers. ### What do I need to run an agentic investigation with bitdrift? You need three things: the bitdrift CLI installed and authenticated via bd auth login, the /bd-cli skill loaded in your agent, and at least one workflow configured in the bitdrift platform for your app. Once a funnel or exploration workflow exists, the agent can discover and query it automatically. ### What does a bitdrift funnel workflow show? A funnel workflow tracks user progression through defined steps in your mobile app, such as Categories, Browse, Cart, and Payment, and reports the conversion rate between each step. The agent can pull seven days of chart data, compare step-level drop-off rates, and correlate them with crash or error data from the same period to identify the likely cause of a conversion decline.