AgentComposite

LLM Configuration

Each agent is bound to an LLM configuration that specifies which model to use, the temperature, and the maximum context window size. This is stored as a LLMConfig value object:

public class LLMConfig
{
    public string Provider { get; set; }       // "Anthropic", "OpenAI", "AzureOpenAI", "Ollama"
    public string Model { get; set; }          // "claude-sonnet-4-6", "gpt-4o", etc.
    public int CredentialId { get; set; }      // resolved via ICredentialResolver
    public float Temperature { get; set; }     // 0.0 = deterministic, 1.0 = creative
    public int MaxContextTokens { get; set; }  // e.g. 200000 for Claude
    public int MaxOutputTokens { get; set; }   // max tokens in one response
    public string? Endpoint { get; set; }      // for Azure OpenAI or local models
}

Memory Configuration

The MemoryConfig on the agent controls which memory types are active and their retention settings:

public class MemoryConfig
{
    public bool EpisodicEnabled { get; set; }      // default: true
    public bool SemanticEnabled { get; set; }      // default: true (if vector store configured)
    public bool ProceduralEnabled { get; set; }    // default: false
    public int EpisodicTtlDays { get; set; }       // 0 = never expire
    public int MaxWorkingMemoryTokens { get; set; } // context window budget
    public WorkingMemoryPruningStrategy PruningStrategy { get; set; } // FIFO, Summarize, SlidingWindow
    public int SemanticTopK { get; set; }          // how many chunks to inject per turn
}

Tool Registry

The ToolRegistry holds all MCP tool definitions that this agent can call. Tools are discovered at agent startup by querying each registered MCP server's GET /tools endpoint:

public class ToolRegistry
{
    private readonly Dictionary<string, ToolDefinition> _tools = new();

    public void Register(ToolDefinition tool) => _tools[tool.Name] = tool;

    public IReadOnlyList<ToolDefinition> GetAll() => _tools.Values.ToList();

    public ToolDefinition? Get(string name) =>
        _tools.TryGetValue(name, out var t) ? t : null;
}

public class ToolDefinition
{
    public string Name { get; set; }
    public string Description { get; set; }        // LLM reads this to decide when to call
    public JsonSchema InputSchema { get; set; }    // JSON Schema for tool parameters
    public JsonSchema? OutputSchema { get; set; }
    public string McpServerUrl { get; set; }       // which server handles this tool
}

Persona

The agent persona defines the user-facing identity:

public class AgentPersona
{
    public string DisplayName { get; set; }     // shown in chat UI
    public string AvatarUrl { get; set; }
    public string Tone { get; set; }            // "formal", "casual", "technical"
    public string Language { get; set; }        // "en", "es", "fr", etc.
    public bool EmojiAllowed { get; set; }
    public string DomainExpertise { get; set; } // "HR", "Finance", "IT Support"
}

Goal Definition

Goals are structured (not free-form) and feed into the system prompt via template variables:

public class AgentGoal
{
    public string Purpose { get; set; }          // "Help employees with HR questions"
    public List<string> Capabilities { get; set; }  // what the agent can do
    public List<string> OutOfScope { get; set; }    // what it should decline
    public List<string> EscalationCriteria { get; set; } // when to escalate to human
}

// In the system prompt template:
// {{goal.Purpose}}
// Capabilities: {{goal.Capabilities | join ', '}}
// Do not assist with: {{goal.OutOfScope | join ', '}}

Agent Lifecycle in Code

Creating and registering an agent in Octopus:

// Registration at startup (agents loaded from database)
services.AddOctopus(options =>
{
    options.UsePlugin<SqlServerStoragePlugin>();
    options.UsePlugin<SemanticKernelPlugin>();
});

// Creating an agent via the service layer
var agent = new AgentComposite
{
    Name = "HR Assistant",
    TenantId = tenantId,
    LLMConfig = new LLMConfig
    {
        Provider = "Anthropic",
        Model = "claude-sonnet-4-6",
        CredentialId = 42,    // stored in credential vault
        MaxContextTokens = 200_000,
        Temperature = 0.3f
    },
    Persona = new AgentPersona
    {
        DisplayName = "Aria",
        Tone = "formal",
        Language = "en"
    }
};

await _agentService.CreateAsync(agent);