AIAgent

MPO Version: 1.3.0

Defines a standalone AI agent at the service level. Each agent encapsulates a model configuration, system prompt, input/output pipeline, optional tool access, and conversation management. Agents support multiple execution modes (task, chat, orchestrator) and modalities (text, image, audio, video, vision). They are exposed as dedicated REST and/or SSE endpoints and can be invoked by other agents as tools.

interface AIAgent = {
  agentBasics : AgentBasics;
  modelSettings : AgentModelSettings;
  inputSettings : AgentInputSettings;
  outputSettings : AgentOutputSettings;
  toolSettings : AgentToolSettings;
  chatSettings : AgentChatSettings;
  orchestrationSettings : AgentOrchestrationSettings;
  imageSettings : AgentImageSettings;
  audioSettings : AgentAudioSettings;
  videoSettings : AgentVideoSettings;
  customEndpoint : AgentCustomEndpoint;
  endpointSettings : AgentEndpointSettings;
  guardrails : AgentGuardrails;
}

AgentBasics

MPO Version: 1.3.0

Core identity and execution mode of the AI agent. The name serves as the asset identifier and is used in endpoint paths, inter-agent references, and code generation.

interface AgentBasics = {
  name : String;
  description : Text;
  executionMode : AgentExecutionMode;
  modality : AgentModality;
}

AgentExecutionMode

Defines how the AI agent processes requests.

const AgentExecutionMode = {
  task: "task",
  chat: "chat",
  orchestrator: "orchestrator",
};

AgentModality

Defines the primary input/output modality of the agent.

const AgentModality = {
  text: "text",
  vision: "vision",
  image: "image",
  audio: "audio",
  video: "video",
};

AgentModelSettings

MPO Version: 1.3.0

AI provider and model configuration for the agent. Defines which LLM or generative model to use, its parameters, and optional dynamic system prompt sources that inject contextual data into the system message at runtime.

interface AgentModelSettings = {
  provider : String;
  model : String;
  systemPrompt : Text;
  temperature : Float;
  maxTokens : Integer;
  responseFormat : AgentResponseFormat;
  systemPromptSources : AgentContextSource[];
}

AgentResponseFormat

Defines the expected response format from the AI model.

const AgentResponseFormat = {
  text: "text",
  json: "json",
  url: "url",
  binary: "binary",
};

AgentContextSource

MPO Version: 1.3.0

Defines a dynamic data source that fetches context at runtime. Used in both systemPromptSources (injected into the system message) and contextSources (injected into the user prompt). Each source produces a named variable accessible via MScript in templates.

interface AgentContextSource = {
  name : String;
  sourceType : AgentContextSourceType;
  config : MScript;
}

AgentContextSourceType

Defines how a context source fetches its data.

const AgentContextSourceType = {
  dataObject: "dataObject",
  libraryFunction: "libraryFunction",
  apiCall: "apiCall",
};

AgentInputSettings

MPO Version: 1.3.0

Defines how incoming request data is transformed into the agent's prompt or input. Includes the prompt template (an MScript expression that builds the user message from request body, params, and context), file upload support for vision/audio agents, and dynamic context sources that fetch additional data to enrich the prompt.

interface AgentInputSettings = {
  promptTemplate : MScript;
  acceptsUpload : Boolean;
  maxFiles : Integer;
  autoResize : Boolean;
  autoConvert : Boolean;
  contextSources : AgentContextSource[];
}

AgentOutputSettings

MPO Version: 1.3.0

Controls how the AI model's response is transformed before returning to the client. Supports extracting a specific property from JSON responses, applying post-processing transformations, and optionally storing generated media files to a permanent bucket.

interface AgentOutputSettings = {
  responseProperty : String;
  postProcess : MScript;
  storeTo : MScript;
}

AgentToolSettings

MPO Version: 1.3.0

Configures which tools the agent can invoke during its execution loop. Only applicable for text and vision modality agents. Tools follow the OpenAI/Anthropic function calling convention. The agent's tool-calling loop continues until the model stops requesting tools or the maxToolCalls guardrail is hit.

interface AgentToolSettings = {
  enableCrudTools : Boolean;
  crudScopes : String;
  apiTools : String;
  agentTools : String;
  libraryTools : String;
  customTools : AgentCustomTool[];
}

AgentCustomTool

MPO Version: 1.3.0

Defines a custom tool that the agent can invoke during its execution loop. Custom tools are used when existing CRUD, API, or library tools are not sufficient. Each tool has a name, description (shown to the AI), a JSON Schema defining its parameters, and a handler pointing to a library function.

interface AgentCustomTool = {
  name : String;
  description : String;
  parameters : Text;
  handler : MScript;
}

AgentChatSettings

MPO Version: 1.3.0

Conversation history management for chat-mode agents. Controls how messages are stored, truncated, and optionally summarized. Only applicable when executionMode is 'chat'.

interface AgentChatSettings = {
  historyStorage : AgentHistoryStorage;
  maxHistoryMessages : Integer;
  summarizeAfter : Integer;
  refreshSystemPrompt : Boolean;
}

AgentHistoryStorage

Defines where conversation history is stored for chat-mode agents.

const AgentHistoryStorage = {
  memory: "memory",
  database: "database",
};

AgentOrchestrationSettings

MPO Version: 1.3.0

Multi-agent orchestration configuration for orchestrator-mode agents. Defines the execution strategy and the steps to coordinate. Only applicable when executionMode is 'orchestrator'.

interface AgentOrchestrationSettings = {
  strategy : AgentOrchestrationStrategy;
  maxIterations : Integer;
  completionCondition : MScript;
  orchestrationSteps : AgentOrchestrationStep[];
}

AgentOrchestrationStrategy

Defines how orchestrator agents execute their steps.

const AgentOrchestrationStrategy = {
  sequential: "sequential",
  parallel: "parallel",
  adaptive: "adaptive",
};

AgentOrchestrationStep

MPO Version: 1.3.0

Defines a single step in an orchestrator agent's workflow. Each step invokes a sub-agent with mapped input and captures its output. Steps can be conditional and can run in parallel with other steps.

interface AgentOrchestrationStep = {
  name : String;
  agentName : String;
  inputMapping : MScript;
  outputMapping : MScript;
  condition : MScript;
  parallel : Boolean;
}

AgentImageSettings

MPO Version: 1.3.0

Image generation parameters. Only applicable when modality is 'image'. These settings are passed to the provider's image generation API (e.g., DALL-E, Fal AI).

interface AgentImageSettings = {
  size : String;
  quality : AgentImageQuality;
  style : AgentImageStyle;
  count : Integer;
}

AgentImageQuality

Defines the quality level for image generation.

const AgentImageQuality = {
  standard: "standard",
  hd: "hd",
};

AgentImageStyle

Defines the artistic style for image generation.

const AgentImageStyle = {
  natural: "natural",
  vivid: "vivid",
};

AgentAudioSettings

MPO Version: 1.3.0

Audio generation or transcription parameters. Only applicable when modality is 'audio'. Supports text-to-speech (TTS) and speech-to-text (STT) directions.

interface AgentAudioSettings = {
  direction : AgentAudioDirection;
  voice : String;
  outputFormat : AgentAudioOutputFormat;
  language : String;
}

AgentAudioDirection

Defines whether the audio agent performs text-to-speech or speech-to-text.

const AgentAudioDirection = {
  tts: "tts",
  stt: "stt",
};

AgentAudioOutputFormat

Defines the output format for text-to-speech audio generation.

const AgentAudioOutputFormat = {
  mp3: "mp3",
  wav: "wav",
  opus: "opus",
  flac: "flac",
};

AgentVideoSettings

MPO Version: 1.3.0

Video generation parameters. Only applicable when modality is 'video'. Video generation is typically asynchronous — the agent polls for completion and streams progress events via SSE.

interface AgentVideoSettings = {
  duration : Integer;
  aspectRatio : String;
  resolution : String;
}

AgentCustomEndpoint

MPO Version: 1.3.0

Custom provider endpoint configuration. Only applicable when provider is 'custom'. Allows connecting to any OpenAI-compatible API (self-hosted models, vLLM, Ollama, etc.) or Anthropic-compatible endpoints.

interface AgentCustomEndpoint = {
  baseUrl : String;
  apiKeyEnvVar : String;
  apiFormat : AgentApiFormat;
}

AgentApiFormat

Defines the API format for custom provider endpoints.

const AgentApiFormat = {
  openai: "openai",
  anthropic: "anthropic",
  raw: "raw",
};

AgentEndpointSettings

MPO Version: 1.3.0

Controls how the agent is exposed as an API endpoint. Each agent can have a REST endpoint (synchronous) and/or an SSE endpoint (streaming). Authentication requirements and path customization are configured here.

interface AgentEndpointSettings = {
  hasRestEndpoint : Boolean;
  hasSseEndpoint : Boolean;
  basePath : String;
  authRequired : Boolean;
}

AgentGuardrails

MPO Version: 1.3.0

Safety limits and cost controls for the agent. Prevents runaway costs, excessive tool usage, and enforces input/output validation. These limits are enforced at runtime by the agent execution loop.

interface AgentGuardrails = {
  maxToolCalls : Integer;
  maxTokenBudget : Integer;
  maxFileSizeMb : Integer;
  timeout : Integer;
  inputValidation : MScript;
  outputValidation : MScript;
  allowedMimeTypes : String;
}