3A Studio

Artificial. Agentic. Architecture.

IDE for visual multi-agent systems. Build, manage, and monitor AI agent flows with a drag-and-drop editor, prompt evaluation, knowledge base, and standards enforcement -- backed by SQLite (local dev) / PostgreSQL Neon (Vercel deploy).

Active repo: https://github.com/stsgs1980/AAA-studio (348 commits) Canonical source: https://github.com/stsgs1980/3a-studio-mas (45K LOC, full project) Frozen repo: https://github.com/stsgs1980/3a-studio (143 commits, stopped 30.05.2026)

Table of Contents

• Features
• Tech Stack
• Getting Started
• Architecture
• Screens
• Monorepo Packages
• Agent Typology
• Resource Map
• Entity Relationships
• Configuration
• Development Rules
• Documentation
• Deployment
• License

Features

• 19 screens: Dashboard, Flow Editor, Templates, Agents, Agent Creator, Hierarchy, Pipelines, Workflows, Prompt Studio, Knowledge Base, Skill Forge, Standards Manager, Audit Log, Settings, Tasks, Testing, Quality Analyzer, Self-Correction, Wiki
• Flow Editor with 20 node types (5 AI + 4 Mgmt + 4 Data + 2 Knowledge + 2 Integration + 3 Special) and ReactFlow v12 live execution
• Prompt Studio with 6 modules: Write (live scoring), Formulas (10), Frameworks (11), Techniques (17), Compare, Intent
• Knowledge Base with TF-IDF semantic search and document upload
• Skill Forge with CRUD, code/tests, StandardsPicker, and SKILL.md export
• 10 agent typology patterns based on cross-framework research (LangChain, CrewAI, AutoGen, Anthropic, OpenAI Agents SDK, Google ADK, Amazon Bedrock)
• Agent Creator with guided wizard for creating new agents
• Agent Hierarchy with 7 edge types (command, sync, twin, delegate, feedback, supervise, broadcast)
• Standards Manager with rules editor and cross-reference validation
• Monorepo with 5 packages: @stsgs/ui, @stsgs/prompting, @stsgs/shared, eslint-plugin-3a, verify-docs
• i18n support (EN/RU, 7 namespaces) with Landing page, Auth (login/signup/verify/reset/forgot)

Tech Stack

• Framework - Next.js 15/16 (App Router), React 19
• Language - TypeScript 5
• Styling - Tailwind CSS 4 + midnight theme (#0D1117 base, #58A6FF accent)
• Database - Prisma ORM + SQLite (local dev) / PostgreSQL Neon (Vercel deploy)
• State - Zustand
• Flow Editor - @xyflow/react (React Flow v12)
• Auth - jose JWT + Edge middleware + httpOnly cookies
• LLM - z-ai-web-dev-sdk (multi-provider with mock fallback)
• Code Highlighting - shiki (github-dark)
• Animations - Framer Motion

Getting Started

Prerequisites

• Node.js 20+ or Bun

Installation

git clone https://github.com/stsgs1980/AAA-studio.git
cd AAA-studio
cp .env.example .env
bun install
bun run db:push

Run

bun run dev

Default login: admin / admin.

Architecture

3A Studio replaces 3 repos x 110 skills x manual sync with one database.

StsDev-Wiki (solutions, ADR)
        |
3A Studio (living system)
  Standards Manager = coding standards (was FRONTEND_STANDARD.md)
  Skill Forge       = skill registry (was 110 folders)
  Prompt Studio     = prompt engineering (was prompting_sts/)
  Flow Editor       = visual orchestration (20 node types, ReactFlow)
  Knowledge Base    = document search (was ChromaDB, now SQLite/PostgreSQL)
  Audit Log         = activity tracking (was session-log)
        |
Zai-agent-toolkit (compiled skills export for Z.ai sandbox)

Changed a standard -- reflected everywhere. Created a skill -- assign to agent -- wire into flow -- run pipeline.

Screens

Screen	Route	Description
Dashboard	/dashboard	KPIs, sparklines, heatmap, timeline -- live from DB, auto-refresh 30s
Flow Editor	/editor	20 node types (5 AI + 4 Mgmt + 4 Data + 2 Knowledge + 2 Integration + 3 Special), ReactFlow, live execution, per-node model + usage tracking
Templates	/templates	6 flow templates + prompt library, clone to editor
Agents	/agents	CRUD, executions, Skills/Standards EntityPicker, system prompt
Agent Creator	/agent-creator	Guided agent creation wizard
Hierarchy	/hierarchy	Visual parent/child agent graph, 7 edge types (command/sync/twin/delegate/feedback/supervise/broadcast)
Pipelines	/pipelines	Real flow execution, node-level drill-down
Workflows	/workflows	Workflow management and monitoring
Prompt Studio	/prompt-studio	6 modules: Write (live scoring) + Formulas (10) + Frameworks (11) + Techniques (17) + Compare + Intent
Knowledge Base	/knowledge	Upload, TF-IDF semantic search
Skill Forge	/skills-page	CRUD, code/tests, StandardsPicker, SKILL.md export
Standards Manager	/standards	CRUD, rules editor, cross-ref validation
Audit Log	/audit	JSON-highlighted details, filter by entity
Settings	/settings	Multi-provider LLM, theme/language, key masking
Tasks	/tasks	Task tracking and management
Testing	/testing	Test runner, judge scoring, metrics
Quality Analyzer	/quality-analyzer	Quality analysis and scoring
Self-Correction	/self-correction	Auto-revision loop
Wiki	/wiki	16 articles, Ctrl+K drawer

Additional: Landing page (/), Auth (login/signup/verify/reset/forgot), i18n (EN/RU, 7 namespaces).

Note: Analysis and Comparison screens have API routes. Approvals (HITL), Cost Monitor, and Testing screens are implemented (Wave 7.5).

Monorepo Packages

Package	Purpose
@stsgs/ui	Design tokens (Midnight palette), ThemeProvider, cn utility
@stsgs/prompting	6-criteria scoring, 10 formulas, 11 frameworks, 17 techniques, intent detection, comparison
@stsgs/shared	Core types: Agent, Skill, Standard, Flow, Knowledge, Prompt, Audit
eslint-plugin-3a	4 rules: max-lines (150), max-use-state (3), no-cross-layer, no-unicode-escapes
verify-docs	Data-driven README consistency checker with cross-repo validation and plugin system

Agent Typology

Based on cross-framework research (LangChain, CrewAI, AutoGen, Anthropic, OpenAI Agents SDK, Google ADK, Amazon Bedrock).

#	Pattern	Core Idea	Phase
1	Tool-Calling	Native function calling, default agent	Phase 3
2	Router	Classify input, route to specialist	Phase 3
3	Specialist (CrewAI)	Role + Goal + Backstory persona	Phase 3
4	Orchestrator + Workers	Break task, delegate, synthesize	Phase 3
5	Evaluator	Generate, score, feedback loop	Phase 3
6	ReAct	Thought/Action/Observation loop	Phase 4
7	Plan-and-Execute	Plan once, execute steps, re-plan	Phase 4
8	Autonomous	Open-ended loop with tools, stop conditions	Phase 5
9	Parallel/Voting	Multiple LLM calls, aggregate	Phase 5
10	Prompt Chaining	Sequential LLM calls with gates	Phase 4

Phase 3 focuses on types 1-5. These cover the majority of real-world multi-agent scenarios.

Resource Map

The strategic resource inventory that feeds 3A Studio.

#	Donor	LOC	Status	What to borrow
1	3a-studio-mas	~45K	SOURCE	Flow Editor (20 nodes), Prisma Schema (37 models), Prompting System, LLM Client, Diagnostics, Pipeline Middleware, Dashboard, Auth, Monorepo packages
2	P-MAS_init	54,798	ARCHIVED	Workflow execution engine, ReactFlow hierarchy v2, Workflow Pipeline UI, 7 edge types, Resilience layer, WebSocket service
3	MVP-Flow-Studio-Pro	18,193	ARCHIVED	6 advanced nodes (Switch/Merge/Loop/Webhook/Variable/DataSource), 26 multi-agent templates, Template Gallery, i18n
4	P-MAS-architector	~81,000	ACTIVE	Orchestrator, prompt versioning, citation system, executor pipeline, 67 skills
5	prompting-v0.0	4,304	PACKAGE	Pure TS prompting lib: 20 techniques, 11 frameworks, 6-dim scoring, 20 formulas, 12 orchestration patterns
6	Flow-Studio-Pro	~750	ACTIVE	5 unique nodes (Loop, Delay, Merge, SubAgent, Search)

Warning: In 3a-studio-mas, src/lib/prompting/ contains stubs (random scoring). Always use packages/prompting/ which has real heuristic implementations.

Entity Relationships

Standard <--(standardIds)--> Skill <--(skills)--> Agent --> Flow --> PipelineExecution
Standard <----(standards)--- Agent
Agent --> AgentExecution
Agent --> AgentHierarchy (parent/child)
Flow --> FlowVersion
KnowledgeCollection --> KnowledgeDocument

All cross-entity references validated on delete (409 Conflict if referenced).

Configuration

Variable	Required	Description
DATABASE_URL	Yes	SQLite (local) or PostgreSQL Neon (Vercel) connection
DIRECT_URL	Vercel	PostgreSQL direct connection (migrations)
AUTH_SECRET	Yes	JWT signing secret (64-char hex)
ENCRYPTION_KEY	Yes	AES-256-GCM key for API key encryption (64-char hex)
ADMIN_USERNAME	No	Demo login username (default: admin)
ADMIN_PASSWORD	No	Demo login password (default: admin)
ZAI_API_KEY	No	Z.ai SDK API key
ZAI_BASE_URL	No	Z.ai SDK base URL

Development Rules

1. Files <= 150 lines (eslint-plugin-3a/max-lines)
2. Max 3 useState per component (eslint-plugin-3a/max-use-state)
3. No cross-layer imports (eslint-plugin-3a/no-cross-layer)
4. Midnight palette -- consistent dark theme across all screens
5. Zustand for all shared state
6. Prisma + SQLite/Neon -- single source of truth

Documentation

Doc	Description
WORKFLOW.md	Full workflow, architecture diagram, roadmap phases

Deployment

Deployed to Vercel via GitHub auto-deploy (push to main).

License

TBD

Agent Rules

Any AI agent working on this project MUST read and follow AGENT_RULES.md before performing any operations.

Configuration

See AGENT_RULES.md for project-specific agent configuration, commands, and architecture.

---

Built with: Next.js 15.3 + React 19 + TypeScript 5.8 + Prisma 6.8.2 + Neon PostgreSQL + Socket.IO + Zustand + Radix UI