Learning module · Intermediate
Deploy a narrow agent: shadow mode and guardrails
Give one agent one job end to end, inside the tools your team already uses, and prove it in shadow mode before it goes live.
Watch
Why it matters
What this protects
Most companies still struggle to show measurable value from AI in production. The agents that work do one job end to end, inside the tools the team already uses, so nobody has to learn new software.
Rani Haddad's firm, Ghidorah, builds these for small and mid-sized businesses. For a hospitality staffing company, one agent keeps a live operations dashboard and runs worker onboarding over WhatsApp and email through to the signed contract, saving about 160 hours a month. For a medical-supplies distributor, an agent in the sales team's WhatsApp group turns each order into an invoice and any purchase orders, and the back office approves them.
Language models are probabilistic. Keep each agent narrow, give it only the tools its job needs, and let code handle anything that must come out the same every time.
Game plan
Do it in 7 steps
- Spend a week tracing how work moves through the business, then pick the most accessible task an agent could own end to end.
- Build around the tools the team already uses.
- Keep the agent narrow: a tight instruction, only the tools the job needs, often no internet.
- Write code for anything that must be exact every time, such as a daily report, and have the agent run it.
- Run the agent in shadow mode for two weeks to two months, checking every output.
- Go live with one human approval step, and keep tuning it as models change.
- Measure what it does: reports, ERP actions, emails sent, customer conversations handled.
Prompts
Copy and use
Replace the [bracketed] parts. Each prompt keeps you in control: nothing is sent, deleted or shared without your approval.
Any business chat assistant.
Map one process in my business: who does it, which tools they use, how often, and how long it takes. Then pick the single task an agent could own end to end with one human approval step, and describe how we would run it in shadow mode for two weeks before going live.
The system prompt of the agent.
You do one job: [job]. You can use only [tools]. If someone asks about anything else, say it is outside your job and stop. Don't browse the web. When a step must come out the same every time, such as building the daily report, run [approved script] instead of writing new code.
An agent with read access to the chat and your inventory.
Here are last week's messages from [our sales group chat]. Pull out every order: customer, product, quantity. Check each against [inventory export] and draft the invoices and any purchase orders we need. Draft only. Our back office approves every invoice before it goes out.
Check your work
You're done when
- You've named one task, its owner and how you'll measure it.
- The agent has run in shadow mode and its output has been checked.
Taught live in the AI Practicum on Thursday, October 15, 2026. The videos above are clips from that class.
Keep going