Agents in the cloud

Claude Code on your laptop is an agent with you watching. The moment you close the lid, it stops. The leverage you actually want is the colleague who starts at 7 AM whether or not you are awake: reads what came in overnight, scores it, drafts the reply, and puts a digest in your inbox.

Most people imagine this requires infrastructure they cannot run. It does not. It is one scheduled function on the Vercel project you already have, calling the Claude API with the same role file you wrote in Protocol 03, writing to the same Supabase tables. Same folder idea, different home.

Three kinds of agent

The folder is the same in all three. What changes is where it runs and who is watching.

• Claude Code: an agent that writes code
  Runs on your laptop with you in the loop. You give it a plan, it builds, you read the diff. This is the agent you have used for two days. Close the lid and it stops.
• Routines: agents that run on your laptop
  Claude Routines, GrokBot and the like. The same folder on a schedule or a trigger, on your machine: every morning at seven, read the inbox and draft. No server. It stops when the laptop sleeps, which is fine for a daily routine and wrong for anything a customer waits on.
• Server agents: Claude remote agents and Vercel functions
  The same role definition, running in the cloud with nobody watching. A Vercel cron wakes it, it reads your database, thinks with Claude, writes its notes and its log. Always on, costs cents, and this is the one you build this afternoon.

The five parts

Every cloud agent you will ever build has these five parts. Name them before you build and the build is short.

• Trigger
  A cron schedule in vercel.json (every morning at 7, converted to UTC) or a webhook (a form was submitted). Protected by a secret so nobody else can fire it.
• Context
  The rows it reads from Supabase. Inquiries created in the last 24 hours. Anything stuck in new_lead for more than a week. Never the whole database; exactly what the job needs.
• Instructions
  The role file at agents/<name>/ROLE.md. Who it is, your hot-lead rules, the forbidden list, and the exact JSON it must return per row: a score, a one-sentence reason, a suggested next step, a draft reply.
• Tools
  What it may do: write a note, write a draft, send you a digest. What it may never do: email a customer, delete anything, change a stage past contacted. Least privilege. Widen when you trust it.
• Log
  agent_runs (when it ran, how many rows, status, error) and agent_notes (every decision, every draft, whether you approved it). If it is not in the log, it did not happen.

Where the key lives

The Claude API key is an environment variable on Vercel and in your local .env.local, and nowhere else. Not in code. Not sent to the browser. Not pasted into a prompt. Search the repo for the key prefix before you merge: zero hits, every time.

Cost, rhythm, and the human in the loop

A daily agent that reads fifty rows and writes fifty notes costs cents. The rhythm matters more than the model: once a day at 7 AM is a colleague, every minute is a bill. Start daily. Speed up only when the log shows it would have mattered.

The agent drafts. The admin page shows the draft. You approve. That single button is the difference between a helpful colleague and a liability. Your first cloud agent keeps it. So does your tenth.

An agent on a schedule in the cloud, reading your CRM, drafting replies you approve, logging every decision, with a digest in your inbox every morning. Built from the same folder idea as every agent on your laptop.