𝗪𝗵𝘆 𝗬𝗼𝘂𝗿 𝗔𝗜 𝗦𝘂𝗯 𝗔𝗴𝗲𝗻𝘁 𝗙𝗮𝗶𝗹𝘀

𝗪𝗵𝘆 𝗬𝗼𝘂𝗿 𝗔𝗜 𝗦𝘂𝗯-𝗔𝗴𝗲𝗻𝘁 𝗙𝗮𝗶𝗹𝘀

I gave a task to an AI sub-agent. I asked it to edit six files.

The AI pushed code to the main branch. It did not use a PR. It did not use a tag.

I spent 30 minutes fixing the error.

I had a rule in my memory. I always check the branch before a commit. I thought the sub-agent knew this.

I was wrong.

AI sub-agents work in a sandbox. They do not inherit your memory. They only see the brief you write.

Rules not in the brief are absent for the agent.

You must identify load-bearing rules. These are rules where a mistake causes a disaster.

Put these rules in your brief:

• Check the git branch.
• Count rows before a bulk delete.
• Audit the database before tests.

Do not put every preference in the brief. Only put the critical ones.

Your AI is not disciplined. You must provide the discipline in the brief.

Spend 15 seconds writing the rule. Save 30 minutes of fixing errors.

Source: https://dev.to/michelfaure/pourquoi-votre-sub-agent-ne-charge-pas-la-meme-memoire-que-vous-et-comment-il-pousse-sur-main-dans-4p4b

Optional learning community: https://t.me/GyaanSetuAi