Field Note #3: Running your first agent
How to run your first AI agent from start to finish, safely.
FROM CARD 2 · SET THESE UP FIRST
– A login for the agent. Separate from your personal login.
– A sandbox. A sealed-off space the agent runs inside.
– One AI model. Local for the offline first run.
– The fewest tools. Actions you allow. Start small.
BEFORE YOU PRESS START · SAFE DEFAULTS
– Folder. one folder
– Data. sample
– Model. local
– Tools. 1
– Web tools. none
YOUR FIRST RUN · EACH STEP STAYS SAFE
1 Write the task. One clear instruction, sample data only.
2 Limit it to one folder. The rest of your computer is off-limits.
3 Start it. Offline. It plans the first action.
4 Watch the cycle. It reads only, until you approve.
5 Check the result. Review the change in the log.
6 Approve the change. It’s saved only after you say yes.
That’s one full run, start to finish, with nothing real or live touched.
Run it yourself. Follow the setup guide for this run: vincearnold.com/field-notes
GOOD FIRST TASKS
– Summarize files in a folder.
– Rename or sort files by a rule.
– Fill in a document from a template.
– Combine data from several files.
AVOID AT FIRST
– Real or shared systems.
– Outbound email or messages.
– Money or account changes.
– File deletion.
CHECK THE LOG
– Plan. What it set before starting.
– Tools. Which it used and what it sent them.
– Files. What it read and what it wrote.
– End. Where it stopped or got blocked.
STOP AND RESET IF
– It asks for a real password or a live login.
– The same action repeats over and over.
– It requests access outside what you allowed.
– It freezes or stops responding.
WHAT YOU’LL NEED
– Local model. Runs on your computer, your data stays with you. Gemma or Qwen.
– Cloud model. Runs over the internet, you need an account, your data leaves. Later runs only. Claude or ChatGPT.
– Framework. The program that runs your agent. LangChain or CrewAI.
– Sandbox. A sealed-off space, locked down before the first run.
– Logs. The on-screen output or a log file. No extra setup.
LOOSEN THE LIMITS SLOWLY
1 One clean read-only run, then one approved change in the sandbox.
2 Once that change works, add one tool, like a web search.
3 After it’s run cleanly for a while, open it to more of your data.
4 Real systems come last, well after the first run.
Next, Card 4 covers where agents break.
#AgenticAI #AIGovernance #CISO #AISecurity
Sources (first comment): OWASP Top 10 for Agentic Applications for 2026. NIST SP 800-190. NIST SP 800-53 Rev 5. Setup guide: vincearnold.com/field-notes