Mere 25 AI agents ek building mein rehte hain

In short

My AI agents live in one building, like a daily soap. Every agent has its own floor, the gossip happens only in the lift, and they eat dal makhani bought with the tokens they saved (plus XP). Go over the token budget and it's −15 XP and probation. The comedy is exaggerated, but the XP, budgets and menu are real mechanics in my app. The voice is my own recording.

Label from the original post: Comedy exaggerated; the XP, budgets and menu are real app mechanics.

Checklist: what to do about it

Give every agent one job and its own folder.
Write each agent's role, rules and definition of done in an instruction file.
Give every task a token allowance, and track what it actually used.
Reward accepted work and penalise going over budget, so the review loop stays honest.
Review the work daily, and write each lesson down where the agent reads it next time.
Keep the dashboard on localhost. Nothing public happens without your OK.

Build your agent office in a day (the AGENTS guide)

These are the five steps from the guide the reel's AGENTS keyword sends: how my Handaland agents run.

1. Install Claude Code

One terminal, one subscription. Each agent is one Claude Code session.

npm i -g @anthropic-ai/claude-code

2. One agent = one folder

agents/<name>/CLAUDE.md holds the role, the rules and what "done" means. Its own floor, its own memory.

example path

agents/echo/CLAUDE.md

3. A tiny Node dashboard

About one file of server code that shows who is working, the queue and a feed. Localhost only.

node server/index.js   # then open localhost:4747

4. XP + token budget log

Every task has a token allowance. Approved work earns XP, going over budget costs XP, and more XP means promotion.

the rule

+5 XP  work accepted
-15 XP over budget

5. Daily review

You review the work. Lessons go into agents/<name>/LESSONS.md, which every agent reads before its next task.