The three things that make an agent
Chatbot vs agent, one task
Task: "handle the unpaid invoices." A chatbot returns tips on writing payment reminders. An agent queries your invoicing tool, finds the three that are overdue, drafts a chase email for each in your tone, and pauses: "ready to send these?" One gave you homework; one handed you finished work. (Sound familiar? It's the whole idea behind Yasmine.)
What agents are genuinely good at — and not
Good: multi-step digital chores with a clear goal — reporting, triage, drafting, research, code changes. Bad: tasks needing taste-based final calls, anything you can't verify, and work you'd never delegate to a sharp intern without review. The honest framing is delegation, not replacement — which is why the next section matters more than the demo.
The checklist before you trust one
- Can it ask before acting? An agent without approval gates is autonomy without a brake. Look for per-action approval.
- Where does it run? Shared sandbox with other customers, or an isolated environment per workspace?
- Whose model account? Your own keys mean your data flows under your provider's terms — resold tokens mean a middleman.
- Can you see what it did? Every action should leave a visible trail you can audit.
- Where does memory live, and can you wipe it?
Yasmine is our answer to that checklist: an agent that lives in Slack, per-channel memory, per-tool approval, a dedicated microVM per workspace, on your own keys or Claude account. Here's what that looks like day to day.
