Hi everyone,
I’m working on an idea and I’d like to hear what people in the field think.
The idea: an agent that controls Packet Tracer for you. You give it a document or just write a message like “build 2 routers, a switch and 2 PCs with static routing”. Then it does the work:
reads your document or message
places the devices, connects the cables and configures everything
tests the network with pings and tells you what passes or fails
fixes errors and gives you a report of what it did
You stay in control. You can watch each step, pause it, undo, or ask it to change something.
My main question: right now the agent can use an external AI service (Gemini, Groq...) to understand free-form messages and unfamiliar documents. I’d like to avoid depending on an external API, because of cost, quotas, internet access, and privacy. So:
- is there a way to avoid using them?
- OR Is it better to use rules and templates for the common labs and keep AI only as an optional fallback?
And more generally:
Would you use something like this? For what?
What’s missing, or what would worry you?
Any idea or criticism is welcome. Thanks for reading!
Top comments (2)
Rules and templates sound like a sensible first version for common labs. I’d put a small, explicit topology specification between the prompt and the actions: devices, interfaces, links, addressing and intended reachability. Show that plan for approval before changing the lab, and ask for clarification when a required detail is missing.
For testing, include an expected failure too: PC A can reach the server, guest PC B cannot. That makes a useful acceptance case beyond “all pings passed.” An optional model could help translate unfamiliar wording into the same specification, while the validation and execution stay deterministic.
Thanks, that’s very helpful. One question about the plan: when something is missing, like an IP or a gateway, would you rather the agent asks the user right away, or guesses a default and shows it clearly in the plan so the user can correct it?