Very Good Ventures

Had a fantastic time at flutterCon Berlin speaking with partners and friends on the future of Flutter and strategies for continuing to strengthen the community. Very Good Ventures represented with 7 talks! Next stop, #MexicoCityTechWeek. Come say Hi Oct 29 and 30th at the Anthropic Claude Community workshop and the Microsoft AI conference

Before you build an AI agent, ask whether you need one. Agents add cost, latency, and risk. Ana Polo Sánchez, Senior Software Engineer II at VGV, runs through seven questions before she builds one. Does the task end with an action or an answer? For answers, a single model call is usually enough. Does it need data from your own systems? If not, it probably doesn’t need tools. Are the steps the same every time? A fixed sequence is cheaper and easier to test as a plain workflow. Is the input free text or a form? Structured fields rarely need a model to interpret them. Can you write down what it may do? Actions, limits, and escalation cases become hooks. If you can’t list them, the policy work comes first. What is the cost of one wrong action? The higher it is, the more you want identity checks, limits, an audit trail, and a human path. How many actions does it need? One to about five fits a single agent. Beyond that, split the work into subagents. None of them is about model quality, which is the point. The list is one section of a longer post on building an agentic support workflow with Claude and Flutter. It also covers the tool-use loop, hooks, and the Flutter client architecture. Read the full article here: https://lnkd.in/eRKpxDFi

Very Good Ventures has been named a Select Services Partner in the Claude Partner Network. We've been building with Claude, powered by Anthropic, on client work since its launch, and this designation reflects projects that are already live. One of our senior engineers used Claude Code to rebuild Trackhouse Racing's NASCAR telemetry pipeline in Rust in three weeks, so race engineers can now analyze 18 billion data points while the race is still running. For Scooter's Coffee, a multiagent pipeline we built on Claude cut an app migration estimated at six to eight months down to about three or four. Divine went from a rapidly prototyped codebase to a production release in 90 days. Eleven of our engineers are Claude-certified, and we bring AI into every engagement where clients approve it. Teams that aren't ready for AI yet still get the same senior engineering behind their product. Jorge Coca wrote about what this means for teams building with Flutter. The link is in the first comment. Anthropic Claude #Flutter #ClaudeCode #AgenticAI

Running flutter run -d linux will not get your app onto an Automotive Grade Linux screen. That command targets Flutter's GTK desktop embedder. AGL images run their own Wayland compositor and expect a purpose-built embedder to host the app. Jorge Coca, our VP of Engineering, just published the setup behind our new AGL in-vehicle infotainment demo. It runs on three open-source tools from Toyota Connected. meta-flutter bakes the Flutter engine and your app into the AGL image. emb_cli cross-compiles the AOT bundle and deploys it to the device over SSH. ivi-homescreen hosts the app directly on Wayland, with no GTK involved. You don't need a car to start. Jorge builds on a Raspberry Pi 5 with a 10.1" touchscreen. He has tested the 1GB and 8GB models, and both handle most use cases well. You can also start with an AGL emulator on your computer. Once the pipeline is in place, you build the in-vehicle UI with the same Flutter widgets you already use on mobile, web, and desktop. If you're porting an existing app, check your plugins for Linux support and against the ivi-homescreen plugin list before you start. The toolchain isn't limited to cars. NuStep's gym and physical therapy equipment runs on it, as shown at #FlutterCon USA 2026. We joined Automotive Grade Linux in spring 2026, and we'll debut the demo at the AGL All Members Meeting Europe in Berlin. The demo code is open source, so you can start from ours. Full guide and repo in the first comment!

We had a great time at FlutterConf Latam last week, where our team gave two talks! Rémy Baudet spoke about building a cross-functional AI workflow from Figma to Flutter, walking through how AI can tie design and engineering together throughout that handoff. Sofía Rey San Esteban gave a lightning talk on offline AI in Flutter with Gemma, covering the real costs of running an LLM locally with metrics pulled straight from her own demo. Sofia was also on the event staff this year, helping organize the whole conference. Always great to be part of this community. Thanks to everyone who stopped by to chat!

A GenUI dashboard went from 54 seconds to 3.8, at zero generation cost, with the same completion rate. Andrew King plugged TypeSafe AI's Jev into a plant-monitoring demo. Jev is a decision model: you hand it a state and typed questions, it returns typed answers with confidence scores, in parallel. It generates no text and never touches a measured value. What he caught early was the tell. The data was right and the reasoning was right, but the frontier model was making zero tool calls. The data was already fetched. The model was retyping numbers into a template and deciding what goes on screen, one token at a time. That's classification work. Jev answers about 100 closed questions in half a second. Code fetches the data, assembles from a pre-approved component catalog, and fills in every number at render. Most of the 3.8 seconds is the query. The limits are real: synthetic plant data, prose from templates, and anything outside the catalog falls back to an LLM. In a regulated environment that hard ceiling on what can render is a feature. Before swapping models, list every decision your composer makes. If more than half are closed questions with a fixed answer set, you're paying generation prices for classification. Full breakdown with benchmark table at the link in comments! #vgv #jev #genui

A GenUI dashboard went from 54 seconds to 3.8, at zero generation cost, with the same completion rate. Andrew King plugged TypeSafe AI's Jev into a plant-monitoring demo. Jev is a decision model: you hand it a state and typed questions, it returns typed answers with confidence scores, in parallel. It generates no text and never touches a measured value. What he caught early was the tell. The data was right and the reasoning was right, but the frontier model was making zero tool calls. The data was already fetched. The model was retyping numbers into a template and deciding what goes on screen, one token at a time. That's classification work. Jev answers about 100 closed questions in half a second. Code fetches the data, assembles from a pre-approved component catalog, and fills in every number at render. Most of the 3.8 seconds is the query. The limits are real: synthetic plant data, prose from templates, and anything outside the catalog falls back to an LLM. In a regulated environment that hard ceiling on what can render is a feature. Before swapping models, list every decision your composer makes. If more than half are closed questions with a fixed answer set, you're paying generation prices for classification. Full breakdown with benchmark table at the link in comments! #vgv #jev #genui

I'm preparing Season 02 of the #VeryGoodEngineering podcast. I already have a few stories lined up, but also I'd love to hear from my network, as the podcast is a gift for you all! What guests would you like to see sharing their story with me? What topics would you like to hear more about in the podcast? What changes would you like to see in the podcast? Drop a comment, so we can keep making Very Good Engineering the podcast where you learn and have fun, both at the same time! https://lnkd.in/ggNtscyB