Free-form AI that directly executes network commands is dangerous in production, because one hallucinated or misjudged action can take down a live network. Ticvic's network AI agent never executes CLI directly -- it proposes a plan, a deterministic execution layer runs only pre-approved templates, and a human authorizes the change.
We built an AI agent that does not autonomously execute network commands by design.
Here's why that's its best feature.
The Free-Form AI Fantasy
Every week, a new demo claims: "Watch our AI log into a switch and resolve an outage autonomously." Technically impressive? Perhaps. Production-ready? Not without strict controls, governance, and safeguards. Here's what those demos don't show you:
- The AI hallucinating a VLAN ID that doesn't exist.
- The AI misinterpreting "shut" as "shutdown" on the wrong interface.
- The AI executing a config change that violates compliance.
The 3 AM Scenario (That Keeps Us Up at Night)
Imagine this: Critical outage. 3 AM. A junior engineer, desperate to restore service, prompts the AI: "Fix the connection to the core router." A free-form AI, lacking business context, executes a script that:
- Flaps the BGP session (acceptable)
- Also reloads the entire routing table (not acceptable)
- Also disables the ACL for your largest customer's financial data (catastrophic)
Outage fixed? Yes. Compliance breach started? Also, yes. This isn't hypothetical. It's what happens when you let AI touch CLI without guardrails.
The "Ban AI Entirely" Trap
Most organizations see this risk and do the sensible thing: they ban AI entirely. Now you're back to:
- Manual SSH sessions at 3 AM.
- Engineers memorizing arcane CLI syntax.
- On-call burnout.
- Cognitive overload masking the real signals.
Banning AI isn't a strategy. It's surrender to the status quo.
There Is a Third Way
At Ticvic, we asked a different question: what if AI never touched the CLI at all? What if AI did what it's actually good at — reasoning, pattern recognition, risk analysis — and left the execution to deterministic, policy-bound systems? That question became our architecture.
How It Works (The 10,000-Foot View)
1. Intent Detection (Rule-Based, Not AI)
AI never guesses what you want. We map natural language to predefined intents using rules. If it doesn't match, it fails.
2. Entity Extraction (Reusable Validators)
Extractors pull out IPs, interface names, VLAN IDs. No assumptions. Missing a subnet mask? HTTP 400. No defaults.
3. Execution Plan (Deterministic)
The planner outputs a step-by-step plan before anything executes. No plan = no execution. Full stop.
4. Validation Layer (The Guardrails)
Every action is validated against business rules. VLAN creation requires a valid VLAN ID, a valid IPv4 address, and a valid subnet mask. Missing any? Rejected.
5. Execution Engine (Netmiko)
Only now does the CLI get touched. And it's not AI doing the touching — it's deterministic Python code executing predefined commands.
6. Parser Layer (Dedicated, Not Generic)
Each SHOW command has its own parser. No generic regex hell. No brittle parsing. Structured JSON guaranteed.
7. Summarization (AI on Facts Only)
Finally, AI gets to summarize. But it sees only verified facts — grouped interfaces, counts, names. No raw CLI. No guessing.
The Result
- AI does what AI is good at: reasoning and summarization
- Deterministic code does what deterministic code is good at: execution
- The two never mix
AI is never trusted for execution. All execution paths are predefined. Nothing executes unless a valid plan exists. This isn't anti-AI. It's pro-production.
See this running in production — SD-WAN reference architecture