crestapps-core-ai-agents

CrestApps.Core AI Agents

An agent is an AIProfile with Type = AIProfileType.Agent. The primary model sees each available agent as a tool that takes one prompt argument and returns the agent's answer. Each agent has its own system prompt, deployment and settings, so instructions do not collide.

Guidelines

• AddCoreAIOrchestration() (or AddAISuite(...)) registers AgentToolRegistryProvider; nothing else is needed.
• Every agent needs a Name and a strong Description. Agents with an empty name or description are skipped, and the description is all the model (or an MCP client) has to decide with.
• Use OnDemand for most agents and link them per profile. AlwaysAvailable agents are added to every request and cost tokens on every turn.
• By default an agent runs as one completion with tools disabled. Set AgentMetadata.AllowToolInvocation = true only when it needs its own tools; it then runs through its orchestrator, and agents inside it are still suppressed.
• Leave ChatDeploymentName empty to inherit the caller's chat deployment, or name a cheaper or more capable model for the agent.

Create an Agent

using CrestApps.Core.AI.Models;
using CrestApps.Core.AI.Profiles;

var agent = await profileManager.NewAsync("translator");

agent.Type = AIProfileType.Agent;
agent.DisplayText = "Translator";
agent.Description = "Translates text between languages. Provide the target language and the text.";
agent.ChatDeploymentName = "gpt-5-mini";

agent.Put(new AIProfileMetadata { SystemMessage = "Translate faithfully. Return only the translation." });
agent.Put(new AgentMetadata { Availability = AgentAvailability.OnDemand });

await profileManager.CreateAsync(agent);

Link Agents to a Chat Profile

chatProfile.Put(new AgentInvocationMetadata
{
    Names = ["translator", "code-reviewer"],
});

await profileManager.UpdateAsync(chatProfile);

The names flow to AICompletionContext.AgentNames; AgentToolRegistryProvider returns those agents plus every AlwaysAvailable agent as ToolRegistryEntrySource.Agent entries.