From Silicon to PWA: My Experience Deploying an Autonomous AI Agent (Hermes) Directly into My Daily Workflow

Yacov drori — Thu, 18 Jun 2026 20:44:34 +0000

As an engineer with 20+ years of experience working from the silicon up—programming bare-metal C, motion control, and industrial automation—I’ve always been skeptical of "AI coding assistants." Writing auto-complete snippets is easy; architecting complex, production-ready systems is a completely different beast.

But recently, during the launch of my new ventures—specifically a modular Flutter Web PWA Book-to-App Framework (tefila_meditation), a Go-native desktop app (remote-viewer), and optimization of a women's health cycle calculator—I decided to do something different.

I deployed Hermes (by Nous Research), an autonomous, CLI-driven AI agent, directly into my Linux terminal environment.

Unlike standard chat widgets that sit in a browser tab and ask you to copy-paste code, Hermes was deployed locally with full filesystem, package manager, and terminal access. It doesn't write "stubs" or "placeholders"; it writes code, compiles it, analyzes the static output, runs git, and deploys the builds directly.

Here are 3 real-world battles we won together:

The Mobile Autoplay & PWA Asset Path Hell
When building our offline-ready Daily Quiet Corner playing looping ambient nature audio, we hit a classic mobile web wall. Modern browsers aggressively block audio unless triggered by an immediate, synchronous user gesture. Furthermore, Flutter Web compiles assets into a double-nested structure (assets/assets/audio/...), causing a triple-asset path resolution (assets/assets/assets/...) and silent 404s.
• How we solved it: Hermes analyzed the network resource requests of our test browser, isolated the 404s, and refactored our player into a synchronous execution flow to preserve the browser's "user gesture token." It dynamically switched the source loader to use native UrlSource with double-asset routing on Web, while preserving native AssetSource on mobile.
Timezone-Drift & DST Calculations in Python/Dart
Our cycle calculations were suffering from a classic bug: seasonal Daylight Saving Time (DST) transitions were introducing 23-hour and 25-hour day calculations, shifting predictions.
• How we solved it: Instead of applying patchy offset fixes, we normalized all datetime math to UTC across both the Dart frontend and Python backend, backed by comprehensive unit tests to ensure 100% mathematical consistency during seasonal clock switches.
Desktop Linux Toolchain Compilation
When building our Go-native desktop app "remote-viewer" on Ubuntu, we ran into native GStreamer link errors during CMake. Hermes parsed the compiler logs, isolated the missing native dev packages, and provided the exact GStreamer installation dependencies to unblock Android Studio.

Key Takeaways:
• Say Goodbye to Async Gaps in UI Events: Mobile browsers are ruthless with autoplay blocks. If you have even a single 'await' microtask between a click and a '.play()' call, the browser discards the gesture token. Keep your media play setups strictly synchronous.
• Flutter Web Asset Caching is Aggressive: Service workers in PWAs cache Compiled JS so heavily that standard refreshes won't show updates. Always perform a hard cache-reload (Ctrl+F5) when testing asset-prefix changes.
• AI Agents as Partners, Not Replacements: The magic happens when you combine systems-thinking (the human) with rapid execution and iteration (the agent). I guided the architecture; Hermes handled the boilerplate, compilations, and repo synchronization.

Deploying Hermes directly into my terminal didn’t make me code less; it allowed me to build faster, cleaner, and with much higher confidence.

Are you integrating AI agents into your development workflows, or keeping them in the sandbox? Let’s discuss!

DEV Community: Yacov drori

From Silicon to PWA: My Experience Deploying an Autonomous AI Agent (Hermes) Directly into My Daily Workflow