A few months ago I finished setting up my Raspberry Pi with metrics, alerting, domain access, and secrets, and ended up with a pretty neat device that did nothing. Around the same time OpenClaw started showing up in my feed — a bot with access to an LLM that could sit on a small device and do work for you. The pitch was good enough that I gave it a weekend.
A weekend turned into months of configurations hell and confusion. The thing I underestimated was how big the gap is between "an LLM bot responds on Discord" and "an LLM bot does real engineering work for me." A bot that can chat is easy. A bot that can take a Discord message, decide it needs to write code, write that code, run tests, open a PR, and tell me about it that is a different category of system.
An LLM bot does real engineering work for me..
That shift surfaces three problems that the cosy "ask the LLM a question" framing hides. Given my experience in building agents, I wanted to avoid the context rots and instruction overload problems from the start. Additionally, I had been pretty comfortable using Claude Code for coding tasks. So OpenClaw had to come-in for something else.
If you have are agents builder, you will recognise that building a multi agent system is the answer to the problems above. The killer feature of OpenClaw for me is the chat integrations which comes at a low cost so it felt really compelling.
