Stop giving your agent a PhD when it needs a checklist

I see it every week. Someone spends three days setting up vector embeddings, RAG pipelines, and multi-model routing for an agent that doesn't know how to check their email.

You're solving the wrong problem. Your agent doesn't need a knowledge graph — it needs clear instructions and a place to write things down.

Here's the setup that actually works:

The 3-file agent: AGENTS.md for instructions, MEMORY.md for persistence, one cron job for scheduling. That's it.

Start with AGENTS.md. This is your agent's job description:

# AGENTS.md

## Daily Tasks
- Check email for urgent items (keywords: "urgent", "asap", "emergency")
- Update project status in Notion
- Send morning briefing by 9 AM

## When to escalate
- Budget questions over $500
- Customer complaints mentioning "cancel"
- Any security-related emails

## Available tools
- Gmail API (read/send)
- Notion API (read/write)
- Slack webhook for notifications

Notice what's missing? No "use your best judgment." No "be helpful." Just specific tasks, clear escalation rules, and a list of what the agent can actually do.

Next, MEMORY.md. This is where your agent writes down what it learns:

# MEMORY.md

## Important contacts
- Sarah (sarah@company.com): Prefers morning emails, handles billing
- Mike (mike@vendor.com): Always CC his manager on urgent requests

## Patterns I've noticed
- Deploy emails usually come Friday afternoons
- Customer complaints spike after feature releases
- Budget approvals take 2-3 days in Q4

Your agent appends to this file every time it learns something useful. No vector database, no embeddings — just a text file that gets included in the context window.

The cron job ties it together:

# Every 30 minutes
30 * * * * /usr/bin/python3 /path/to/agent.py --check-email

# Daily briefing at 8:30 AM
30 8 * * * /usr/bin/python3 /path/to/agent.py --morning-briefing

# Weekly memory cleanup on Sundays
0 10 * * 0 /usr/bin/python3 /path/to/agent.py --clean-memory

This setup handles 80% of what people want from an agent:

• Consistent execution of routine tasks
• Learning from interactions over time
• Clear boundaries on what it can and can't do
• Reliable scheduling without manual intervention

But here's the key: you can deploy this today. No infrastructure, no database migrations, no complex prompt engineering. Just three files and a cron job.

Warning: This approach won't impress anyone at a conference. It won't scale to enterprise complexity. But it will work reliably while you figure out what your agent actually needs to do.

Start here. Add complexity only when you hit real limitations, not imaginary ones. Your agent needs to crawl before it can run — and most agents never need to run at all.

The best agent system is the one that's actually running in production, handling real work, making your life easier. Everything else is just expensive procrastination.