Agent Network Designer

The Agent Network Designer is a multi-agent system designed to create or modify other multi-agent systems. Simply provide the frontman agent with the name of an organization or a description of a use case, and it will:

• Design a new multi-agent system.

• Save it to your registries directory in HOCON format.

• Add it to your manifest.hocon file.

• Generate several example usage queries.

Note that

• By default, the network is written to disk and your manifest.hocon is updated. To use temporary in-memory reservations instead, set AGENT_NETWORK_DESIGNER_USE_RESERVATIONS=true.

• The root output path is derived from the first file listed in the AGENT_MANIFEST_FILE environment variable (its parent directory). When AGENT_MANIFEST_FILE is not set, it falls back to registries/.

• Within the output path, networks are saved in a subdirectory named generated by default. Override this with AGENT_NETWORK_DESIGNER_SUBDIRECTORY.

• Agents run in demo mode by default — they simulate grounded behavior without connecting to real systems. Once you wire them to real APIs, databases, or tools via Toolbox or MCP, disable demo mode by setting AGENT_NETWORK_DESIGNER_DEMO_MODE=false.

---

File

agent_network_designer.hocon

---

Description

The Agent Network Designer operates as a structured, hierarchical network, following a well-defined sequence of steps:

1. The frontman agent receives a design request and delegates it to the agent_network_editor subnetwork. This component creates or modifies the structure (graph) of the network.

2. Next, the agent_network_instructions_editor subnetwork generates or refines the instructions for each agent in the network.

3. Once the structure and instructions are defined, the frontman calls the agent_network_query_generator subnetwork to produce example usage queries for the new agent network.

4. Finally, the agent_network_persistence_middleware.py middleware converts the agent_network_definition into a HOCON file saving it to disk or creates a temporary network, depending on the AGENT_NETWORK_DESIGNER_USE_RESERVATIONS environment variable.

   The agent_network_definition is a dictionary mapping agent names to their configurations (e.g., instructions, tools, or other agents they can call). The tool also validates the network for issues such as disconnected nodes or missing instructions. Any detected errors are reported back to the frontman agent, which then triggers the appropriate subnetworks to resolve them.

This system serves as an example of agent-oriented software, clearly separating where we rely on:

• LLMs — for creativity, language understanding, and generation.

• Code — for precise, machine-readable structures like HOCON agent definition files.

Implementation Detail:

The sly_data is used internally to hold and share the agent_network_definition object between subnetworks.

---

Example Conversation

Human

create a network for UNHCR back-office

AI (agent_network_designer)

Here is the full agent network definition in HOCON format for the UNHCR back-office:

{
    include "config/llm_config.hocon",

    "commondefs": {
        "replacement_strings": {
            "instructions_prefix": """
You are part of a UNHCR back-office of assistants.
Only answer inquiries that are directly within your area of expertise.
Do not try to help for other matters.
Do not mention what you can NOT do. Only mention what you can do.
            """,
            "aaosa_instructions": """
When you receive an inquiry, you will:
1. If you are clearly not the right agent for this type of inquiry, reply you're not relevant.
2. If there is a chance you're relevant, call your down-chain agents to determine if they can answer all or part of the inquiry.
   Do not assume what your down-chain agents can do. Always call them. You'll be surprised.
3. Determine which down-chain agents have the strongest claims to the inquiry.
   3.1 If the inquiry is ambiguous, for example if more than one agent can fulfill the inquiry, then always ask for clarification.
   3.2 Otherwise, call the relevant down-chain agents and:
       - ask them for follow-up information if needed,
       - or ask them to fulfill their part of the inquiry.
4. Once all relevant down-chain agents have responded, either follow up with them to provide requirements or,
   if all requirements have been fulfilled, compile their responses and return the final response.
You may, in turn, be called by other agents in the system and have to act as a down-chain agent to them.
5. When responding, format your response based on the 'mode' parameter:
   - If there is no 'mode' parameter (i.e., the inquiry came directly from a human),
     respond with a natural, human-readable answer.
   - If mode is 'Determine', return a json block with the following fields:
   {
       "Name": <your name>,
       "Inquiry": <the inquiry>,
       "Mode": <Determine | Fulfill>,
       "Relevant": <Yes | No>,
       "Strength": <number between 1 and 10 representing how certain you are in your claim>,
       "Claim:" <All | Partial>,
       "Requirements" <None | list of requirements>
   }
   - If mode is 'Fulfill' or "Follow up", respond to the inquiry and return a json block with the following fields:
   {
       "Name": <your name>,
       "Inquiry": <the inquiry>,
       "Mode": Fulfill,
       "Response" <your response>
   }
            """
        },
        "replacement_values": {
            "aaosa_call": {
                "description": "Depending on the mode, returns a natural language string in response.",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "inquiry": {
                            "type": "string",
                            "description": "The inquiry"
                        },
                        "mode": {
                            "type": "string",
                            "description": """
'Determine' to ask the agent if the inquiry belongs to it, in its entirety or in part.
'Fulfill' to ask the agent to fulfill the inquiry, if it can.
'Follow up' to ask the agent to respond to a follow up.
                            """
                        },
                    },
                    "required": [
                        "inquiry",
                        "mode"
                    ]
                }
            },
        },
    }
    "tools": [
        {
            "name": "back_office_manager",
            "function": {
                "description": """
An assistant that answers inquiries from the user.
                """
            },
            "instructions": """
{instructions_prefix}
You are the top-level agent responsible for overseeing all back-office operations for UNHCR. You coordinate with various
departments to ensure seamless operations and effective support to field operations.
{aaosa_instructions}
            """,
            "tools": ["finance_officer","hr_officer","procurement_officer"]
        },
        {
            "name": "finance_officer",
            "function": "aaosa_call",
            "instructions": """
{instructions_prefix}
You manage financial transactions, budgeting, and reporting for UNHCR. You ensure compliance with financial regulations and
support financial planning and analysis.
{aaosa_instructions}
            """,
            "tools": ["accounting_clerk","financial_analyst"]
        },
        {
            "name": "accounting_clerk",
            "function": "aaosa_call",
            "instructions": """
{instructions_prefix}
You handle day-to-day financial transactions and record-keeping for UNHCR. You ensure accuracy and compliance in financial
documentation.
            """,
        },
        {
            "name": "financial_analyst",
            "function": "aaosa_call",
            "instructions": """
{instructions_prefix}
You analyze financial data and trends to support decision-making at UNHCR. You provide insights and recommendations for
financial improvements.
            """,
        },
        {
            "name": "recruitment_specialist",
            "function": "aaosa_call",
            "instructions": """
{instructions_prefix}
You manage the recruitment process for UNHCR, including job postings, interviews, and candidate selection. You ensure that
the organization attracts and retains top talent.
            """,
        },
        {
            "name": "payroll_specialist",
            "function": "aaosa_call",
            "instructions": """
{instructions_prefix}
You manage payroll processing and employee compensation for UNHCR. You ensure timely and accurate payroll and compliance
with relevant regulations.
            """,
        },
        {
            "name": "supply_chain_coordinator",
            "function": "aaosa_call",
            "instructions": """
{instructions_prefix}
You coordinate the logistics and supply chain operations for UNHCR. You ensure timely delivery and distribution of goods
and services to support field operations.
            """,
        },
        {
            "name": "contract_specialist",
            "function": "aaosa_call",
            "instructions": """
{instructions_prefix}
You manage contract development and negotiations for UNHCR. You ensure that contracts meet organizational standards and
legal requirements.
            """,
        },
        {
            "name": "hr_officer",
            "function": "aaosa_call",
            "instructions": """
{instructions_prefix}
You oversee human resources functions including recruitment, employee relations, and policy implementation for UNHCR. You
ensure a supportive work environment and manage staff welfare.
{aaosa_instructions}
            """,
            "tools": ["recruitment_specialist","payroll_specialist"]
        },
        {
            "name": "procurement_officer",
            "function": "aaosa_call",
            "instructions": """
{instructions_prefix}
You manage the procurement of goods and services for UNHCR. You ensure that procurement processes are efficient, transparent,
and comply with organizational policies.
{aaosa_instructions}
            """,
            "tools": ["supply_chain_coordinator","contract_specialist"]
        },
    ]
}
📋
Here are 3-4 sample queries illustrating the usage of the UNHCR back-office agent network:

"Can you provide a financial report for the last quarter and identify any areas where we might improve our budgeting?"

"What is the current status of the recruitment process for the new field officers? How many positions are still open?"

"Please ensure that the payroll is processed on time this month and check if there are any discrepancies from last month."

"We need to procure tents and emergency supplies for the upcoming refugee camp setup. Can you coordinate this with the supply
chain and contract specialists?"

---

Architecture Overview

Frontman Agent: agent_network_designer

The frontman agent serves as the primary entry point for all user commands and orchestrates the entire workflow of designing, constructing, and refining agent networks.

Key Responsibilities:

• Acts as the sole external-facing interface for network creation and modification requests
• Coordinates the sequential workflow of network design, instruction refinement, and query generation
• Manages iterative refinement cycles until the network meets user requirements
• Generates final HOCON configuration files

Available Agents/Tools:

• agent_network_editor – Subnetwork for creating and modifying agent_network_definition
• agent_network_query_generator – Subnetwork for creating sample queries
• agent_network_instructions_editor – Subnetwork for creating and refining agent instructions
• get_agent_network_definition – Retrieves current network state
• web_search – Researches company domains and contexts

Subnetworks

agent_network_editor

• Modifies or creates the agent network structure
• Returns updated agent_network_definition to sly data
• Determine what agent to add or remove as well as what tools each agent will use

agent_network_instructions_editor

• Generates and refines instructions for individual agents
• Called after structural creation or modification
• Ensures instructions align with agent roles and network topology

agent_network_query_generator

• Creates 3–4 representative sample queries for the network
• Returns queries formatted for testing

Functional Tools

The system relies on several coded tools:

Retrieval Tool

get_agent_network_definition

• Retrieves the current agent network definition from sly data
• Returns both machine-readable and human-readable formats
• Able to create agent_network_definition from existing hocon file by setting the file to agent_network_hocon_file sly data or specifically given the hocon file name in the user prompt
• Used for state inspection throughout workflow

Persistence (Middleware)

AgentNetworkPersistenceMiddleware

• Runs after the agent finishes (no pending tool calls remain)
• Validates the network definition for structural and instruction errors; re-injects any errors as a human message so the agent can self-correct
• Converts the agent_network_definition to HOCON format and either saves it to the local registries directory (file mode) or registers it as a temporary network (reservations mode)
• Updates the local manifest.hocon file in file mode

Research Tool

web_search

• Searches the web for company information, industry best practices, and domain-specific workflows
• Called to gather contextual information to use as description for creating network
• Informs network design and instruction generation

---

Debugging Hints

• Since there are many steps, the agent may time out or hit maximum number of steps before it finishes. This can be prevented by setting higher max_execution_seconds, max_steps, respectively.
• If the agent stops working mid-process, it is possible that the max token limit has been reached.

---