Skip to main content

​
Multi-Agent System Design

Scriptonia uses specialized AI agents driven by LLM calls (OpenRouter). Two modes:
  1. Sequential (stages 1–3): One agent per stage — IdeaProcessor → IntentAnalyzer → PlatformSelector.
  2. Parallel (execution): Multiple execution agents (ProductManager, FrontendDev, BackendDev, DevOpsEngineer, ToolManager) generate files in parallel based on platform templates.
Orchestration is implemented in lib/openai.ts (executeX402Workflow) and lib/platform-templates.ts (generatePlatformDeliverables). The X402 workflow and A2A helpers in lib/x402.ts and lib/a2a.ts support coordination and future extensions.

​
Agent Roles and Responsibilities

Defined in lib/x402.ts as AGENT_ROLES:
AgentStageResponsibility
CEOidea_inputEntry; routes to X402Agent and other agents (conceptual in current flow).
X402AgentexecutionOrchestrator for X402 execution, payment, platform.
IdeaProcessoridea_inputAnalyzes idea → analysis, clarifying_questions.
IntentAnalyzerintent_analysisRefines idea + answers → refined_requirements.
PlatformSelectorplatform_selectionMaps requirements → platform_options, recommended.
PaymentValidatorpayment_processingValidates wallet/balance (in lib/solana and payment API).
ProductManagerexecutionPRD, PLAN, requirements; coordinates with Frontend/Backend, ToolManager.
FrontendDevexecutionUI, pages, components; coordinates with Backend, ToolManager.
BackendDevexecutionAPIs, DB, services; coordinates with Frontend, AIEngineer, ToolManager.
AIEngineerexecutionAI pipelines, RAG, prompts (in AGENT_ROLES; execution file assignment may omit).
DevOpsEngineerexecutionInfra, CI/CD, deployment; coordinates with ToolManager.
ToolManagerexecutionGitHub, meta, export, external services; coordinates with all.

​
Communication Patterns

​
Sequential (Idea → Intent → Platform)

  • IdeaProcessor: Input: idea. Output: analysis, clarifying_questions. No inter-agent call.
  • IntentAnalyzer: Input: idea, answers. Output: refined_requirements. No inter-agent call.
  • PlatformSelector: Input: refined_requirements. Output: platform_options, recommended. No inter-agent call.
Data flows via WorkflowContext (DB or in-memory). Each step is a single LLM request.

​
Parallel (Execution)

  • File–agent mapping: getAgentForFile in lib/platform-templates.ts:
    • PRD, PLAN, requirements → ProductManager
    • frontend, components, pages → FrontendDev
    • backend, api, database → BackendDev
    • infra, deployment, ci_cd → DevOpsEngineer
    • github, meta, export → ToolManager
  • Coordination: generateFileContentWithAgent receives existingDeliverables and workflowContext (idea, analysis, answers, platform, payment). Agent-specific system prompts and “related files” from other agents are included so output stays consistent.
  • Parallelism: Files are grouped by agent; agents run in parallel (Promise.all). Within an agent, files are processed in batches (e.g. 5) with small delays to reduce duplicate content and respect rate limits.

​
Coordination Mechanisms

  1. WorkflowContext — Single source of truth for idea, analysis, answers, refinedRequirements, platformOptions, selectedPlatform, paymentDetails, supabaseIntegration. Passed into generatePlatformDeliverables and generateFileContentWithAgent.
  2. Platform templates — getWebsiteTemplate, getWebAppTemplate, getAndroidTemplate, getIOSTemplate define folderStructure and techStack. generatePlatformDeliverables walks the tree, assigns each file to an agent, and runs generation.
  3. Agent system prompts — In generateFileContentWithAgent, each execution agent has a short system prompt (e.g. “Coordinate with FrontendDev and BackendDev”) so generated specs and contracts stay aligned.
  4. A2A (optional) — lib/a2a.ts provides sendA2ARequest, sendA2AResponse, shareA2AData, sendA2ANotification, and A2AWorkflowIntegration. Used for tests and can be used for richer agent-to-agent calls during execution.
  5. Review step — reviewDeliverables in openai.ts checks structure and completeness and sets review_passed / review_issues on the execution result.
  6. Deduplication — deduplicateContent regenerates content for duplicate deliverables to improve uniqueness.