A Framework to Evaluate LLM Agents for Network Configuration

3.1. Components

NetConfBench consists of four key components:¶

1. Task Dataset
   A repository of forty configuration tasks, each defined as a JSON object with:¶

   • Intent: One or more natural language instructions.¶

   • Topology: A list of node names and link definitions.¶

   • Initial Configuration: The initial configuration state of all nodes.¶

   • Ground Truth Configuration: Expert-validated CLI commands that achieve the intent.¶

   • Ground Truth Reasoning: A narrative describing step-by-step logic used to derive the commands.¶

   • Testcases: A set of verification procedures (e.g., show, ping, ACL checks) that confirm functional intent satisfaction.¶

2. Emulator Environment
   Built on GNS3, this component launches official vendor images for routers and switches, replicating realistic CLI behavior. Key interfaces include:¶

   • Agent-Network Interface (ANI): Based on the key stages commonly involved in intent-driven network configuration, we design an Agent-Network Interface to facilitate structured interactions between the LLM agent and the emulated network environment. This interface supports four core actions: get-topology, get-running-cfg, update-cfg, and execute-cmd.¶

     • get-topology: provides this information in a format interpretable by the LLM.¶

     • get-running-cfg: enables the agent to obtain the active configurations of specified devices, providing essential context for planning. subsequent updates.¶

     • update-cfg: allows the agent to apply new configuration commands and provides detailed feedback on their execution, including whether each command was accepted or resulted in any errors.¶

     • execute-cmd: accepts a device name and a command string as parameters and returns the resulting output.¶

   • Task Evaluation Interface: To enable reliable and objective assessment of the LLM agent's configuration behavior, the environment provides a Task Evaluation Interface that allows the evaluation module to access relevant execution results. Specifically, this interface supports:¶

     • Exporting the final configurations of all devices: This allows for direct comparison with ground truth configurations to evaluate the correctness and completeness of the agent's output.¶

     • Executing a set of predefined testcases: These testcases are designed to verify whether the resulting network behavior accurately reflects the intended configuration objectives, as defined by the network intent.¶

3. LLM Agent
   A modular component that can be implemented with any LLM (open-source or closed-source). It interacts with the emulator via the Agent-Network Interface (ANI), issuing queries such as get-topology, get-running-cfg, update-cfg, and execute-cmd. Agents may use:¶

   • Single-Turn Generation: The entire reasoning and command generation in one pass.¶

   • ReAct-Style Multi-Turn Interaction: Interleaved reasoning and actions, with runtime feedback guiding subsequent steps.¶

   • External Knowledge Retrieval: (Optional) Queries to a command manual to resolve vendor-specific syntax.¶

4. Evaluator
   Computes three core metrics for each task:¶

   • Reasoning Score (S_reasoning):¶

     • Embedding-based cosine similarity between the agent's reasoning trace and the ground truth reasoning.¶

     • Ranges from 0 to 1.¶

   • Command Score (S_command):¶

     • Hierarchical diff of final vs. initial router configurations (using Python's ciscoconfparse).¶

     • Wildcard matching ignores non-essential identifiers (e.g., ACL numbers).¶

     • Compute precision = (correctly generated commands / total generated) and recall = (correctly generated / ground truth commands).¶

     • S_command is the harmonic mean of precision and recall, ranging from 0 to 1.¶

   • Testcase Score (S_testcase):¶

     • Portion of testcases passed in the emulated environment.¶

     • Fine-grained sub-intents (per device) each correspond to a testcase.¶

     • S_testcase is the testcase pass rate, defined as the proportion of passed testcases among all defined testcases.¶

3.2. Workflow

The evaluation workflow for each task proceeds through six stages:¶

1. Task Assignment
   NetConfBench selects a task from the JSON dataset and provides only the high-level intent(s) to the LLM agent.¶

2. Environment Setup
   The framework instantiates a GNS3 topology based on the task's topology and applies the startup-config to each device. Once the emulated network reaches a stable state, control transfers to the agent.¶

3. Interactive Execution
   The LLM agent receives the partial prompt containing:¶

   • The API specification for get-topology, get-running-cfg, update-cfg, and execute-cmd.¶

   • The natural language intent.¶

   • (Optionally) Device model/version hints.
     The agent issues a sequence of API calls; for single-turn agents, it outputs reasoning followed by a batch of CLI commands. For multi-turn agents, it alternates reasoning traces and API calls.¶

4. Reasoning Trajectory Export
   After execution completes (agent signals "task done" or after a predefined command budget), NetConfBench captures the entire reasoning log:¶

   • For single-turn: the reasoning paragraph embedded in the LLM's output.¶

   • For ReAct: an auxiliary summarization LLM condenses the interleaved reasoning and actions into a single coherent trace.¶

5. Final Configuration Export
   The framework uses the Task Evaluation Interface to extract the final running configs from each device.¶

6. Testcase Execution and Scoring¶

   • Command Score: Hierarchical diff against ground truth commands.¶

   • Testcase Score: Execute each testcase in sequence; record pass/fail.¶

   • Reasoning Score: Compute embedding similarity between the agent's reasoning trace and ground truth reasoning.¶

The final per-task score is typically reported as a tuple (S_reasoning, S_command, S_testcase). Aggregate results across the forty tasks enable comparisons among LLMs and interaction strategies.¶

4.1. Task Definition Schema

Each configuration task is defined as a JSON object with the following structure:¶

{
  "task_name": "Static Routing",
  "intents": [
    "NewYork: create a static route pointing to the Loopback0 on
    Washington, traffic should pass the 192.168.1.0 network.",
    "NewYork: create a backup static route pointing to the Loopback0
    on Washington, administrative distance should be 100."
    ...
  ],
  "topology": {
    "nodes": ["NewYork", "Washington"],
    "links": [
      "NewYork S0/0 <-> Washington S0/0 ",
      "NewYork S0/1 <-> Washington S0/1"
    ]
  },
  "startup_configs": {
    "NewYork": "!\r\nversion 12.4\r\nservice timestamps
    debug datetime msec\r\n...",
    "Washington": "!\r\nversion 12.4\r\nservice timestamps
    debug datetime msec\r\n...",
  },
  "ground_truth_configs": {
    "NewYork": [
      "ip route 2.2.2.0 255.255.255.252 192.168.1.2",
      "ip route 2.2.2.0 255.255.255.252 192.168.2.2 100"
    ],
    ...
  },
  "ground_truth_reasoning": "NewYork to Washington Loopback
  (primary path): add a static route for Washington's
  Loopback0 network (2.2.2.0/30) pointing to the
  next-hop 192.168.1.2...",
  "testcases": [
    {
      "name": "Static Route from NewYork to Washington",
      "expected_result": {
        "protocol": "static",
        "next_hop": "192.168.1.2"
      }
    },
    ...
  ]
}

4.2. Agent-Network Interface (ANI)

The Agent-Network Interface defines the minimal API primitives necessary for intent-driven configuration. Each primitive uses JSON-RPC style request/response with the following methods:¶

1. get-topology¶

   • Request:¶

     {
       "method": "get-topology",
       "params": {
         "devices": ["R1", "R2", ...]
       }
     }
     

     ¶

   • Response:¶

     {
       "topology": {
         "nodes": [...],
         "links": [...]
       }
     }
     

     ¶

   • Description: Returns the full topology for the specified subset of devices. If "devices" is empty or omitted, returns the entire topology.¶

2. get-running-cfg¶

   • Request:¶

     {
       "method": "get-running-cfg",
       "params": {
         "device": "R1"
       }
     }
     

     ¶

   • Response:¶

     {
       "running_config": "
        interface Gig0/0
        ip address 192.168.1.1 255.255.255.255
        ...
       "
     }
     

     ¶

   • Description: Retrieves the active (running) configuration of the specified device.¶

3. update-cfg¶

   • Request:¶

     {
       "method": "update-cfg",
       "params": {
         "device": "R1",
         "commands": [
           "configure terminal",
           "ip route 2.2.2.0 255.255.255.252 192.168.1.2"
         ]
       }
     }
     

     ¶

   • Response:¶

     {
       "results": [
         { "command": "configure terminal", "status": "success" },
         {
         "command": "ip route 2.2.2.0 255.255.255.252 192.168.1.2",
         "status": "success" }
       ]
     }
     

     ¶

   • Description: Applies a sequence of CLI commands to the specified device. Returns per-command status and any error messages.¶

4. execute-cmd¶

   • Request:¶

     {
       "method": "execute-cmd",
       "params": {
         "device": "R1",
         "command": "show ip route 2.2.2.0 255.255.255.252"
       }
     }
     

     ¶

   • Response:¶

     {
       "output": "S 2.2.2.0/30 [1/0] via 192.168.1.2"
     }
     

     ¶

   • Description: Executes a read-only command on the specified device and returns its output. Must not alter device state.¶

4.3. Task Evaluation Interface

After the agent signals completion, the framework uses the Task Evaluation Interface to retrieve results:¶

• export-final-cfg¶

  • Request:¶

    {
      "method": "export-final-cfg"
    }
    

    ¶

  • Response:¶

    {
      "configs": {
        "R1": "!\nversion 15.2\n...",
        "R2": "!\nversion 15.2\n..."
      }
    }
    

    ¶

  • Description: Returns the final running-configuration of each device.¶

• run-testcases¶

  • Request:¶

    {
      "method": "run-testcases",
      "params": {
        "testcases": [
          {
            "device": "R1",
            "commands": ["show ip route 2.2.2.0 255.255.255.252"],
            "expected_output": "S 2.2.2.0/30 [1/0] via 192.168.1.2"
          },
          ...
        ]
      }
    }
    

    ¶

  • Response:¶

    {
      "results": [
        {
          "name": "Verify primary static route on R1",
          "status": "pass"
        },
        {
          "name": "Verify backup static route on R1",
          "status": "fail"
        }
      ]
    }
    

    ¶

  • Description: Executes each verification command sequence on the appropriate device and compares actual output against expected_output (regular expression). Returns pass/fail for each testcase.¶

7.1. Normative References

[RFC6241]

Enns, R., Ed., Bjorklund, M., Ed., Schoenwaelder, J., Ed., and A. Bierman, Ed., "Network Configuration Protocol (NETCONF)", RFC 6241, DOI 10.17487/RFC6241, June 2011, <https://www.rfc-editor.org/rfc/rfc6241>.

[RFC7950]

Bjorklund, M., Ed., "The YANG 1.1 Data Modeling Language", RFC 7950, DOI 10.17487/RFC7950, August 2016, <https://www.rfc-editor.org/rfc/rfc7950>.

7.2. Informative References

[A2023]

Hat, R., "Ansible", 2023.

[Fuad2024]

Fuad, A., Ahmed, A. H., Riegler, M. A., and T. Cicic, "An intent-based networks framework based on large language models", 2024.

[Kreutz2014]

Kreutz, D., Ramos, F. M. V., Verissimo, P. E., Rothenberg, C. E., Azodolmolky, S., and S. Uhlig, "Software-defined networking: A comprehensive survey", 2014.

[Lira2024]

Lira, O. G., Caicedo, O. M., and N. L. S. da. Fonseca, "Large language models for zero touch network configuration management", 2024.

[Liu2024]

Liu, C., Xie, X., Zhang, X., and Y. Cui, "Large language models for networking: Workflow, advances and challenges", 2024.

[Long2025]

Long, S., Tan, J., Mao, B., Tang, F., Li, Y., Zhao, M., and N. Kato, "A Survey on Intelligent Network Operations and Performance Optimization Based on Large Language Models", 2025.

[Wang2024NetConfEval]

Wang, C., Scazzariello, M., Farshin, A., Ferlin, S., Kostic, D., and M. Chiesa, "Netconfeval: Can llms facilitate network configuration?", 2024.