The practical difference becomes clear in troubleshooting workflows. When asked to triage a ServiceNow ticket, the agent reads the ticket content, gathers context about entities mentioned from the digital twin, automatically performs path traces for connectivity issues, and returns a diagnosis. The complete workflow remains visible to operators throughout the process.
This differs from simple natural language queries where the system translates a question and returns an answer. The agent is building and executing a plan that may involve multiple data sources and analysis steps.
Custom framework for context control
Forward Networks built its own agentic framework rather than adopting existing tools like LangChain or CrewAI. The decision centered on maintaining precise control over context engineering, which Handigol described as the core engineering challenge for agentic AI.
“We built our own because we wanted complete control over how the agent executes,” Handigol said. “The engineering problem mostly comes down to context engineering. How do you define and maintain the context that is necessary for the agent, that you send to the LLM, to get the right answers?”
The team defines context engineering as providing all relevant information without excess noise. Too little information produces wrong answers. Too much information distracts the model from the correct task.
Context is drawn from Forward Network’s hierarchical data stack. At the base layer sits raw configuration, state, and statistics collected directly from devices. The next tier normalizes this raw data into a queryable model showing everything present in the network and its configuration.
