- The agent can still make typo errors.
- He can repeat the mistakes he already made before by deleting my code that solve a previous problem.
- Hallucinations is still exist and never has improved since.
- It mixes 2 languages in 1 app (in my case, local Bahasa Malaysia and English content)
- You may end up doing a lot of work harness the agent, provide a lot of specs.md, instructions.md, skills.md or what ever md.
Same experiences here.
I am using Kimi 2.6 with claude via Ollama inside vscode
ollama launch claude --model kimi-k2.6:cloud
I just open my b4x folder project and execute that via the terminal.
Because the context gets larger each time you prompt and receives a response, one thing I do to avoid hallucinations is after 30 minutes, ask it to summarize our conversations and save it. Then start a new chat and give it the context of the summarized conversations.
After each sussessful iteration, I create a backup of "LastWorkingVersion" so that when it deletes or changes things that already worked, I refer it to that version.
Using AI is a patience exercise because no matter how much you train it, it will lie, it will assume, it will break your code. Can be very frustrating.
Write a function? it does, is the function 100% accurate? Response No. Then it suggests what can be changed and improved. Why not just do it right first time?
All the best guys and girls...