Building a personal JARVIS-style AI automation system

Hey guys, I’m working on building a personal JARVIS-style AI automation system. Not a chatbot, but an assistant that can actually execute real tasks. The idea is to create an AI agent capable of understanding high-level instructions, breaking tasks into logical steps, interacting with websites and web apps, collecting, structuring, and analyzing data, working with spreadsheets and structured storage, generating text and summaries when needed, automating repetitive workflows, running tasks on a schedule or on demand, and keeping context, memory, and execution logs. Planned tech stack: Python as the core logic, browser automation (Playwright), LLM as the decision-making layer, Google Sheets or a local database for memory, and Windows automation for system-level control. The focus is on a modular architecture where each capability is a separate component that can be improved or replaced independently. I’m not selling anything, just building and experimenting. Would appreciate insights from people with real experience in autonomous AI agents, long-running automation systems, task planning with LLMs, detection and safety considerations, and scaling beyond simple scripts. Thanks in advance for any feedback or ideas.

solid stack man. modular is 100% better than monolithic.

regarding detection and running 24/7 , couple things from my exp:

1. detection: out of the box playwright is kinda loud.. anti-fraud systems fingerprints the CDP usage super fast. if u interact with tough sites dont trust just stealth plugins. u gotta look into patching the actual browser binary to kill navigator.webdriver flags at source level. also make sure ur TLS fingerprints (JA3) match the browser.. otherwise python requests leak ur identity.
2. architecture: python scripts leak memory like crazy over time . dont run it as one big loop. wrap the executor part in a docker container and kill it after every task. let the LLM spin up a fresh one each time. keeps fingerprints clean and solves memory issues.
3. storage: google sheets will rate limit u pretty quikcly. local vector db like Chroma or sqlite is much better for context retrieval and wont slow u down.

good luck , separating the planner from the execution is key .

Building a reliable JARVIS-style AI agent in 2026 works best when the LLM acts as a planner and coordinator, not an executor, with modular, defensive automation, layered memory, strict verification loops, and controlled, observable execution rather than free-form autonomy.

In 2026, the most reliable way to build a JARVIS-style AI agent is to use the LLM as a planner and coordinator, not a direct executor—backed by modular, defensive automation, layered memory, strict verification loops, and tightly controlled, observable execution instead of free-form autonomy.

Have you checked Clawd Bot Open Claw? It sounds like what you are describing. Seems to be getting lots of attention recently

lukem121 said:

Have you checked Clawd Bot Open Claw? It sounds like what you are describing. Seems to be getting lots of attention recently

Click to expand...

yesterday ill check it and I like it

Good luck mate. Its looking ambitious, but as they say, if you can dream it, you can do it.