Net Agent Harness

Overview

Net Agent Harness is a personal lab project focused on creating a model-backed harness for network-engineering workflows. It transforms plain-English change requests into structured, staged, and safe network operations.

The core philosophy is a staged, deterministic pipeline—route → plan → render → execute → review—ensuring that LLM reasoning is used for interpretation and planning, while orchestration and validation remain authoritative in Python.

Tech Stack

• Framework — PydanticAI for agentic workflows
• Validation — Pydantic v2 for strict artifact contracts
• Models — Ollama (local), NVIDIA API, and OpenAI
• CLI — Typer and Rich
• Inventory — REST adapters for NetBox and local mock data

Architecture

The system follows an artifact-first, staged pipeline where each stage has clear ownership and hand-off contracts:

• Route — Deterministically routes requests to relevant network domains (VLAN, ACL, etc.)
• Plan — Interprets intent against current state to decide if a change is needed (apply, no_op, or blocked)
• Render — Translates approved plans into vendor-specific API or CLI operations
• Execute — Applies changes behind approval gates (disabled by default in prototype)
• Review — Validates results and records the outcome

Features

• Staged Pipeline — Enforced stages prevent “hallucination-driven” execution
• Deterministic Routing — Prefers Python-based routing logic for safety
• Inventory Grounding — Target resolution is authoritative and source-backed, not assumed by the LLM
• Artifact-First — Every stage produces a durable, typed JSON artifact for traceability
• Multi-Provider — Seamlessly switch between local models (Ollama) and cloud APIs (NVIDIA/OpenAI)

Status

Active prototype. Currently handles end-to-end VLAN change planning with ongoing work to expand domain support for routing, ACLs, and wireless.