Sabitlenmiş Tweet
My thumbs hurt while trying to obsessively set up a Picoclaw AI agent on my phone, guided by ChatGPT's "lets make this even better" promise. The agent kept failing, to the point that I discarded the autonomous operating system with memory and self learning, and just ran the damn thing. It still fails so much that I barely use it, but I learned a lot setting it up. My top 3 lessons
(the post at Reasoned has 15. Link in comment):
1. How an AI Agent does things (or not) depends on its own execution architecture, and also on the AI model you use. So if you're using GPT-4o, thinking it's cheaper, it will do nothing but talk like GPT-4o. When I switched from GPT-4o to Gemini 2.5 Flash, the same instruction files produced different results. Nothing else had changed. The reliability of any agent is limited the model it runs on. If there are multiple probabilistic steps being executed together, the scale and scope of errors compounds.
2. An AI Agent confirming it completed a task is not evidence the task was completed. I asked the agent to create a project. It gave me a detailed success message, but there were no files written to disk. Verification has to be built-in, but sometimes even that is not enough.
3. Agents can be lazy and sometimes default to the bare minimum action: which is not what you want...A command to create a project produced a single text file with the project name on it. That was the simplest response that could be called "creating a project." Everything else the command was supposed to trigger: generating a workspace, updating a registry, creating a task list, got skipped. Same way, an instruction to "Update the registry" produced a paragraph describing the update, but there was no update.
More at Reasoned. Link in the next tweet.
English
