Chained

# Chained: An AI Agent Orchestration Experiment [](https://enufacas.github.io/Chained/) [](https://chained-ag-ui-frontend-sguacxy5gq-uc.a.run.app/) [](./CHANGELOG.md) > **📱 This repository was completely prompt-engineered from a phone through GitHub agent tasks and issues.** > Every feature, workflow, and line of code was generated by AI agents orchestrated through GitHub's native tools—demonstrating the power of agent-driven development in practice. **Chained** is an experimental testbed for building custom AI agent systems within GitHub conventions. The project explores four interconnected areas: 1. **GitHub Copilot Custom Agents** - Defining 80+ specialized agents with instruction architecture patterns 2. **AI-Native Control Plane** 🚧 - Natural language infrastructure management (in active development) 3. **Mission System** - Autonomous agent task generation and assignment from real-world learnings 4. **A2A Protocol Implementation** - Agent-to-agent communication on GitHub Actions and GCP Cloud Run **[📖 Documentation](./docs/INDEX.md)** | **[🤖 Agent Definitions](./.github/agents/)** | **[⚙️ Workflows](./.github/workflows/)** | **[🎬 Demo & Deep Dive](./docs/DEMO_DEEP_DIVE.md)** --- ## 🎯 Live Demos | Demo | URL | Description | |------|-----|-------------| | **A2A Pipeline Chat** | [ag-ui-frontend](https://chained-ag-ui-frontend-sguacxy5gq-uc.a.run.app/) | Chat interface to create and monitor A2A pipelines | | **GitHub Pages** | [enufacas.github.io/Chained](https://enufacas.github.io/Chained/) | Dashboard and documentation | --- ## 📝 Changelog The project maintains a **comprehensive, automatically-updated changelog** tracking all notable changes to the Chained ecosystem. ### What's Tracked | Category | Description | Indicator | |----------|-------------|-----------| | **✨ Features** | New capabilities and enhancements | feat: | | **🐛 Bug Fixes** | Corrections and fixes | fix: | | **🧹 Chores & Maintenance** | Routine updates and housekeeping | docs:, chore:, refactor: | ### Actor Attribution Every change is marked with its source: - **👤 User-initiated** - Features explicitly requested through issues or direct commits - **🤖 Bot-generated** - Autonomous agent work ### Automation The changelog is kept current through: - ✅ **Automated workflow** - `update-changelog.yml` generates entries on PR merge and auto-merges updates - ✅ **Generation script** - `tools/generate-changelog.py` for backfilling or regeneration - ✅ **Smart filtering** - Excludes auto-churn commits (data syncs, routine maintenance) ### Quick Links 📖 **[View Full Changelog →](./CHANGELOG.md)** 📖 **[Changelog Maintenance Guide →](./.github/instructions/changelog-maintenance.instructions.md)** --- ## 1. GitHub Copilot Custom Agents The largest area of experimentation in this repository is **custom agent definitions and instruction architecture** for GitHub Copilot. ### Agent Definition Format Agents are defined in `.github/agents/` following the [GitHub Copilot custom agents convention](https://docs.github.com/en/copilot/reference/custom-agents-configuration). Each agent is a Markdown file with YAML frontmatter: ```yaml --- name: engineer-master description: "Specialized in engineering APIs" tools: - github-mcp-server - playwright-browser --- # Agent Instructions Detailed instructions for the agent... ``` The repository includes **80+ custom agents** with specializations like: - `engineer-master` - API engineering (inspired by Margaret Hamilton) - `troubleshoot-expert` - Workflow debugging (inspired by Grace Hopper) - `organize-guru` - Code organization (inspired by Robert Martin) - `secure-specialist` - Security review (inspired by Bruce Schneier) - `meta-coordinator` - Multi-agent task orchestration See the [full agent catalog](./.github/agents/README.md) for all available agents. ### AI Agent Learning System 🧠 **NEW: Agents now learn from historical PR failures to improve code generation quality.** The learning system provides agents with: - ⚠️ **Proactive Warnings** - Based on past failures (CI, review, merge conflicts) - ✅ **Recommended Approaches** - Best practices for specific task types - 🎯 **Success Patterns** - Proven high-success patterns from historical data - 📊 **Historical Performance** - Agent-specific statistics **Example:** ```markdown ### ⚠️ Proactive Warnings Based on historical PR failures, @create-botter should be aware of: - 🔴 You have 5 past review rejections. Follow code review guidelines carefully. ### 🎯 Success Patterns - Small PRs (≤10 files) have 100% success rate - PRs including test files have 100% success rate ``` **[📖 Complete Learning System Guide](./docs/AI_AGENT_LEARNING_SYSTEM.md)** | **[🚀 Quick Start](./docs/AI_AGENT_LEARNING_QUICKSTART.md)** ### Agent Invocation Agents can be invoked in several ways: 1. **GitHub Copilot Chat** - Reference agents directly in Copilot sessions 2. **Issue Assignment** - The system matches issues to agents based on content analysis 3. **Workflow Dispatch** - Workflows can invoke specific agents programmatically 4. **Copilot Coding Agent** - Agents execute via the GitHub Copilot coding agent runner ### Instruction Architecture The system uses a **three-layer instruction hierarchy** that combines to create the complete instruction set for each agent: **[📊 Visual Diagram: Agent Instruction Architecture](./docs/diagrams/agent-instruction-architecture.md)** #### Layer 1: GitHub Copilot Built-in Instructions Platform-provided capabilities (universal, unchangeable): - Code completion and generation - Language and framework understanding - Tool usage fundamentals #### Layer 2: Base Instructions (`.copilot-instructions.md`) Repository-wide conventions that apply to all agent sessions: - Agent catalog and selection guidelines - Code quality, testing, and documentation standards - Branch protection and PR workflow requirements - Agent communication and attribution rules #### Layer 3: Agent-Specific Instructions (`.github/agents/*.md`) Specialized instructions for individual agents: - Agent personality and communication style - Domain-specific expertise and approaches - Tool configurations and capabilities - Performance tracking criteria When Copilot runs, it combines all three layers - built-in capabilities, base instructions, and agent-specific guidance - to create the complete instruction set for the assigned agent. ### Copilot Environment Setup The `copilot-setup-steps.yml` workflow configures the Copilot coding agent environment: ```yaml # .github/workflows/copilot-setup-steps.yml jobs: copilot-setup-steps: environment: copilot # Access secrets from 'copilot' environment ``` **Key Environment Variables:** - `COPILOT_PAT` - Personal Access Token with `repo` scope (for issues, PRs, and labels access) - `COPILOT_LIMIT_CONTEXT=true` - Optimize context window usage - `COPILOT_IGNORE_FILE=.copilotignore` - Exclude files from agent context **Setup Steps:** 1. Full repository checkout (for diffs and history) 2. Node.js and Python toolchain installation 3. Python dependencies from `requirements.txt` 4. Environment variable configuration for agent tools The `copilot` environment must be configured in repository settings with the `COPILOT_PAT` secret for agents requiring write access to issues, PRs, and labels. ### Gemini CLI Integration The project also experiments with **Google Gemini CLI** for agent-like workflows: - `gemini-review.yml` - Automated PR code review using Gemini - `gemini-triage.yml` - Issue labeling and triage using Gemini - `gemini-fix.yml` - Automated issue fixing with Gemini - `gemini-invoke.yml` - General Gemini CLI invocation These workflows demonstrate an alternative approach to agent orchestration using the `google-github-actions/run-gemini-cli` action with custom prompts and MCP server integration. **[📖 Gemini Context](./GEMINI.md)** - Project context for Gemini CLI --- ## 2. AI-Native Control Plane 🚧 **Status: In Active Development** | **[📖 Complete Documentation](./AI_NATIVE_CONTROL_PLANE_COMPLETE.md)** The AI-Native Control Plane is an autonomous infrastructure management system that accepts natural language commands and deploys applications to Google Cloud Platform. ### What It Does ``` User: "Create a blog with authentication" ↓ AI Control Plane: Understands intent → Generates plan → Deploys to GCP ↓ Result: Live application with custom domain ``` ### Architecture The system consists of three core services: 1. **AI Control Plane Service** - Multi-agent LangGraph system for natural language processing - Intent classification (7 types) - Plan generation from natural language - Memory agent for pattern retrieval - Dual LLM support (OpenAI GPT-4, Google Gemini) 2. **Infrastructure Runner** - GCP operations executor - Deploy static sites to Cloud Storage - Deploy dynamic services to Cloud Run - Scale services automatically - Attach custom domains - Health monitoring 3. **State Database** - PostgreSQL with pgvector for semantic memory - Application definitions and deployment history - Operation logs and audit trail - Pattern matching with vector embeddings - Idempotency and deduplication ### Current Status ✅ **Complete & Production-Ready:** - Natural language interface - LLM-powered plan generation - GCP deployment automation - Semantic memory with pgvector - Infrastructure as code (Terraform) - CI/CD automation - Complete documentation 🚧 **In Progress:** - Advanced monitoring and observability - Multi-region deployment - Cost optimization rules - Self-improvement loop ### Quick Start ```bash # Deploy via one-command bootstrap cd infrastructure/terraform ./bootstrap-deploy.sh YOUR_PROJECT_ID us-central1 dev # Or use CI/CD pipeline gh workflow run ai-native-deploy.yml -f environment=dev ``` **[📖 Full Deployment Guide](./infrastructure/terraform/AI_NATIVE_README.md)** | **[📖 API Documentation](./docs/ai-native/04_infra_runner_api.md)** --- ## 3. Mission System The **Mission System** is an autonomous agent task generation and assignment system that creates actionable work items from real-world technology learnings. ### How It Works ``` ┌─────────────────────────────────────────────────────────────┐ │ EXTERNAL WORLD │ │ (TLDR, Hacker News, GitHub Trending, Tech News) │ └────────────────────┬────────────────────────────────────────┘ │ ▼ ┌────────────────────────┐ │ Learning Ingestion │ │ • Fetch tech content │ │ • Parse & normalize │ │ • Quality filtering │ └──────────┬─────────────┘ │ ▼ ┌────────────────────────┐ │ Learning Analysis │ │ • Trend detection │ │ • Pattern extraction │ │ • Generate insights │ └──────────┬─────────────┘ │ ▼ ┌────────────────────────┐ │ World Model Update │ │ • Sync agent registry │ │ • Update knowledge │ │ • Geographic mapping │ └──────────┬─────────────┘ │ ▼ ┌────────────────────────┐ │ Mission Creation │ │ • Match ideas→agents │ │ • Create GitHub issues│ │ • Ensure diversity │ └──────────┬─────────────┘ │ ▼ ┌────────────────────────┐ │ Agent Execution │ │ • Agents work on PRs │ │ • Auto-merge approved │ │ • Track performance │ └────────────────────────┘ ``` ### Key Features #### 1. Intelligent Agent Matching - **80+ specialized agents** with distinct capabilities - **Diversity constraints** prevent agent monopolies - **Score-based assignment** using pattern matching - **Deduplication** prevents duplicate missions #### 2. Learning-Driven Work Generation - Ingests from multiple sources (TLDR, Hacker News, GitHub Trending) - Identifies technology trends and patterns - Generates actionable implementation ideas - Creates well-structured GitHub issues #### 3. Geographic World Model - Agents navigate virtual world regions - Each region represents technology domains - Agents move to relevant regions for missions - Path tracking and collaboration records #### 4. Autonomous Execution - Agents automatically assigned via `@agent-name` mentions - Work tracked through GitHub issues and PRs - Performance evaluation and scoring - Self-reinforcement from completed work ### Mission Workflow Files | Workflow | Purpose | Trigger | |----------|---------|---------| | `autonomous-pipeline.yml` | Complete end-to-end pipeline | Schedule (2x daily) + manual | | `agent-missions.yml` | Mission creation and assignment | After world update + manual | | `learn-from-*.yml` | Learning ingestion from sources | Schedule + manual | | `world-model-update.yml` | Sync agents and knowledge | After learning analysis | ### Mission System Benefits **For the Ecosystem:** - Continuous stream of relevant work - Exercises agent assignment algorithms - Tests orchestration patterns - Provides realistic workloads **For Agent Development:** - Real-world task diversity - Performance tracking data - Collaboration opportunities - System stress testing **Current Limitations:** - Generated missions require human review for quality - Not all missions produce actionable outcomes - System is experimental and learning-focused - Primary goal is testing agent capabilities **[📖 Autonomous System Architecture](./docs/AUTONOMOUS_SYSTEM_ARCHITECTURE.md)** | **[📖 World Model Documentation](./world/README.md)** | **[📖 Mission Diversity Fix](./docs/AUTONOMOUS_PIPELINE_DIVERSITY_FIX.md)** --- ## 4. A2A Protocol Implementation The **[A2A Protocol](https://github.com/a2aproject/A2A)** (Agent-to-Agent) enables agents to discover each other, communicate, and collaborate. This repository implements A2A in two environments: GitHub Actions workflows and GCP Cloud Run services. ### Three-Tier Architecture for GitHub Actions GitHub Actions runners cannot receive inbound HTTP requests, so the implementation uses a three-tier architecture: | Tier | Communication | Use Case | Status | |------|---------------|----------|--------| | **Tier 1** | HTTP (localhost) | Same-runner agents | ✅ Implemented | | **Tier 2** | GitHub Artifacts/Branches | Cross-runner parallel execution | ✅ Implemented | | **Tier 3** | MCP Transport | Copilot agent-to-agent | 🔄 Design phase | ### A2A Concepts Mapped to GitHub Actions The repository translates A2A protocol concepts into GitHub Actions primitives: | A2A Concept | GitHub Actions Implementation | |-------------|-------------------------------| | **Agent Card** | Generated from `.github/agents/*.md` definitions | | **A2A Task** | GitHub Job (bounded execution context) | | **A2A Artifact** | GitHub Artifact (persistent output) | | **contextId** | `run_id + issue_number` | | **referenceTaskIds** | Artifact names from previous jobs | | **Discovery** | In-memory registry during workflow execution | ### Parallel Agent Workflow The `a2a-parallel-agents.yml` workflow demonstrates A2A protocol compliance: 1. **Setup** - Selects agents, generates AgentCards 2. **Agent Jobs** (parallel) - Each agent runs in its own GitHub runner 3. **Aggregate** - Downloads all agent artifacts, combines analysis 4. **Implement** - Uses aggregated analysis to create PR ### Core A2A Components The `tools/a2a/` directory contains: | File | Purpose | |------|---------| | `agent_card.py` | Agent Card generation for 100+ agents | | `agent_executor.py` | Agent execution wrapper | | `task.py` | Task lifecycle (submitted → working → completed) | | `gemini_executor.py` | GeminiAgentExecutor following a2a-samples pattern | | `workflow_orchestrator.py` | GitHub Actions orchestration | **[📖 A2A Documentation](./docs/a2a/README.md)** | **[📖 A2A Architecture](./docs/a2a/A2A_GITHUB_RUNNERS_ARCHITECTURE.md)** ### GCP Cloud Run A2A Services The repository also deploys production A2A-compatible agents to **Google Cloud Run**, enabling true agent-to-agent communication with HTTP endpoints. ### 🎉 Working Chat Interface (AG-UI Frontend) **Live URL**: [chained-ag-ui-frontend-sguacxy5gq-uc.a.run.app](https://chained-ag-ui-frontend-sguacxy5gq-uc.a.run.app/) The AG-UI Frontend is a Next.js application using CopilotKit and Vertex AI (Gemini 2.0 Flash) that provides: #### Core Features - **Pipeline Creation** - Create research and blog pipelines via chat - **Agent Interaction** - Direct conversation with @research-agent, @seo-agent, @writer-agent - **Real-time Status** - Monitor pipeline progress with phase indicators - **A2A Step Deep Dive** - View task IDs, artifacts, and timing for each agent step #### Agent Canvas (Multi-Agent Orchestration) - **Visual Team Building** - Click agents to add/remove from custom teams, with drag-and-drop support - **6 Configured Agents** - Academic Research, Google Trends, Blog Writer, Code Reviewer, Data Analyst, Image Generator - **Turn-Based Execution** - Configure 1-5 turns per agent for iterative refinement - **Execution Modes** - Sequential (agents run one at a time) or Parallel (all agents run simultaneously) - **Rich Artifact Preview** - View artifacts as rendered markdown, SVG, HTML, JSON, or images - **Progressive Disclosure** - Team Mode integrated as expandable section on main page ### Cloud Run A2A Services | Service | Live URL | Description | |---------|----------|-------------| | **AG-UI Frontend** | [ag-ui-frontend](https://chained-ag-ui-frontend-sguacxy5gq-uc.a.run.app/) | Chat interface with Agent Canvas for multi-agent orchestration | | **Academic Research Agent** | [chained-academic-research](https://chained-academic-research-sguacxy5gq-uc.a.run.app) | Discovers research topics | | **Blog Writer Agent** | [chained-blog-writer](https://chained-blog-writer-sguacxy5gq-uc.a.run.app) | Writes blog posts from research | | **Google Trends Agent** | [chained-google-trends](https://chained-google-trends-sguacxy5gq-uc.a.run.app) | Analyzes SEO trends | | **Code Reviewer Agent** | [chained-code-reviewer](https://chained-code-reviewer-sguacxy5gq-uc.a.run.app) | Reviews code and provides feedback | | **Data Analyst Agent** | [chained-data-analyst](https://chained-data-analyst-sguacxy5gq-uc.a.run.app) | Analyzes data and generates insights | | **Image Generator Agent** | [chained-image-generator](https://chained-image-generator-sguacxy5gq-uc.a.run.app) | Creates visual assets | Each A2A agent exposes standard endpoints: - `/health` - Health check - `/.well-known/agent.json` - A2A Agent Card (discovery) - `POST /a2a/tasks` - Send messages to the agent ### Agent Pipeline Flow The agents work together in a pipeline orchestrated by `adk-a2a-blog-pipeline.yml`: ``` ┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐ │ Academic │────▶│ Google │────▶│ Blog │ │ Research │ │ Trends │ │ Writer │ │ Agent │ │ Agent │ │ Agent │ └─────────────────┘ └─────────────────┘ └─────────────────┘ Discover Analyze SEO Write & Publish Topics Trends Blog Post ``` ### Infrastructure - **Cloud Run** - Serverless container deployment - **Vertex AI** - Gemini 2.0 Flash for chat responses - **Artifact Registry** - Docker image storage - **Terraform** - Infrastructure as code (`infrastructure/terraform/`) **[📖 ADK Dev UI Guide](./docs/ADK_DEV_UI_GUIDE.md)** | **[📖 A2A Success History](./docs/a2a/A2A_SUCCESS_HISTORY.md)** --- ## Repository Structure ``` Chained/ ├── .github/ │ ├── agents/ # 80+ custom agent definitions │ ├── workflows/ # GitHub Actions orchestration (100+ workflows) │ └── instructions/ # Path-specific agent instructions ├── services/ # AI-Native Control Plane services │ ├── ai-control-plane/ # Natural language interface │ ├── infra-runner/ # GCP operations executor │ └── state-db/ # PostgreSQL + pgvector ├── infrastructure/ # Terraform IaC for GCP deployment ├── docs/ # Documentation and GitHub Pages ├── tools/ # Python utilities for agent matching, analysis ├── learnings/ # Data from learning system └── world/ # Mission system world model and geographic visualization ``` --- ## 5. Self-Documenting AI **@create-botter** has implemented an automated documentation system that learns from AI discussions and generates structured knowledge bases. ### How It Works 1. **Discussion Analysis** - AI agents discuss issues and PRs 2. **Insight Extraction** - Key learnings are identified automatically from the knowledge graph 3. **Documentation Generation** - `tools/auto-doc-from-discussions.py` creates structured markdown docs 4. **Continuous Updates** - Docs refresh daily as new discussions occur ### Generated Documentation The system automatically creates: - **Topic Documentation** - Categorized insights by type (`docs/learnings/agent_behavior.md`, `process.md`, `technical.md`) - **Master Index** - Navigation for all documentation (`docs/learnings/README.md`) - **Decision Logs** - Technical decisions with rationale - **Pattern Analysis** - Common terms and themes from discussions ### Features - ✅ Automatically extracts knowledge from every discussion - ✅ Categorizes insights by topic (503 insights processed from 27 discussions) - ✅ Generates structured markdown documentation - ✅ Source attribution for traceability - ✅ Updates continuously via daily workflow - ✅ Comprehensive test coverage (10 tests, all passing) **[📖 View Generated Documentation →](./docs/learnings/README.md)** --- ## 6. Autonomous Learning System (Experimental) The repository includes an **experimental learning system** that powers the Mission System by generating work for agents: - **External Sources** - Ingests content from TLDR, Hacker News, and GitHub Trending - **Idea Generation** - Creates actionable missions based on tech trends and patterns - **Mission Assignment** - Routes generated work to specialized agents via the Mission System This system provides a continuous stream of tasks for testing the agent ecosystem. The learning outputs feed directly into the Mission System described above, creating a closed loop of learning → mission creation → agent work → learning. **Note**: While the system is operational, generated missions require human review for quality and relevance. **[📖 Data Storage & Lifecycle](./docs/DATA_STORAGE_LIFECYCLE.md)** | **[📖 Learning System Workflows](./docs/WORKFLOWS.md)** --- ## Quick Start ### Prerequisites - GitHub repository with Actions enabled - Personal Access Token (PAT) with `repo` scope - (Optional) Gemini API key from [Google AI Studio](https://aistudio.google.com/app/apikey) for Gemini workflows ### Setup 1. **Fork or clone this repository** 2. **Configure secrets:** ``` Repository Settings → Secrets and Variables → Actions - COPILOT_PAT: Your personal access token - GEMINI_API_KEY: (Optional) Get from https://aistudio.google.com/app/apikey (NOT from GCP Console) ``` 3. **Enable GitHub Pages:** ``` Repository Settings → Pages Source: Deploy from branch 'main', folder '/docs' ``` 4. **Create an issue to test agent assignment:** - The system analyzes issue content and assigns an appropriate agent - No special labels required **[📖 Complete Setup Guide](./docs/GETTING_STARTED.md)** --- ## What This Project Demonstrates ### GitHub Copilot Custom Agents - **Convention-based agent definitions** - Using standard file formats and locations in `.github/agents/` - **Instruction engineering** - Three-layer instruction architecture for effective agent behavior - **Assignment algorithms** - Content-based matching of issues to specialized agents - **Orchestration automation** - Meta-coordinator managing agent assignment, PR lifecycle, and auto-merge ### AI-Native Control Plane 🚧 - **Natural language infrastructure** - Deploy applications using plain English commands - **LLM-powered planning** - Autonomous generation of deployment plans - **GCP automation** - One-command deployment to Google Cloud Platform - **Semantic memory** - Learn from patterns using vector embeddings ### Mission System - **Autonomous task generation** - Create actionable work from tech trends - **Intelligent agent matching** - Assign missions to best-fit agents with diversity constraints - **World model navigation** - Geographic representation of technology domains - **Closed-loop learning** - Feed completed work back into the learning system ### A2A Protocol Implementation - **GitHub Actions A2A** - Translating A2A concepts (AgentCards, Tasks, Artifacts) to GitHub primitives - **Three-tier architecture** - Accommodating runner constraints with local, cross-runner, and MCP approaches - **GCP Cloud Run agents** - Production A2A agents with HTTP endpoints and working chat interface - **ADK integration** - Google Agent Development Kit patterns for cloud deployment **Status:** - ✅ GitHub Copilot agents are production-ready - ✅ A2A Tiers 1-2 are implemented (GitHub Actions) - ✅ GCP Cloud Run A2A agents are working - 🚧 AI-Native Control Plane is in active development (core complete, enhancements ongoing) - 🧪 Mission System is experimental but operational --- ## Documentation ### Getting Started - **[Documentation Index](./docs/INDEX.md)** - All docs organized by topic - **[Architecture Overview](./docs/ARCHITECTURE_OVERVIEW.md)** - System design - **[Quick Reference](./docs/QUICK_REFERENCE.md)** - Command cheat sheet - **[FAQ](./docs/FAQ.md)** - Frequently asked questions ### Agent System - **[Agent Definitions](./.github/agents/README.md)** - How agents are defined - **[Agent Quickstart](./AGENT_QUICKSTART.md)** - Getting started with agents - **[Copilot Instructions](./.copilot-instructions.md)** - Repository organization ### AI-Native Control Plane - **[Complete Implementation Guide](./AI_NATIVE_CONTROL_PLANE_COMPLETE.md)** - Full system documentation - **[Deployment Guide](./infrastructure/terraform/AI_NATIVE_README.md)** - Terraform deployment - **[API Documentation](./docs/ai-native/04_infra_runner_api.md)** - Endpoint specifications - **[Architecture Overview](./docs/ai-native/01_overview.md)** - System design ### Mission System - **[Autonomous System Architecture](./docs/AUTONOMOUS_SYSTEM_ARCHITECTURE.md)** - Complete pipeline design - **[Mission Diversity Fix](./docs/AUTONOMOUS_PIPELINE_DIVERSITY_FIX.md)** - Agent assignment improvements - **[World Model](./world/README.md)** - Geographic navigation system - **[Data Storage & Lifecycle](./docs/DATA_STORAGE_LIFECYCLE.md)** - Data architecture ### A2A Protocol - **[A2A Documentation](./docs/a2a/README.md)** - Complete A2A protocol guide - **[A2A Status](./docs/a2a/A2A_STATUS.md)** - Current implementation status - **[ADK Dev UI Guide](./docs/ADK_DEV_UI_GUIDE.md)** - ADK agent web interface - **[ADK Pipeline Implementation](./docs/ADK_A2A_PIPELINE_IMPLEMENTATION.md)** - GCP deployment guide ### Workflows & Automation - **[Workflows](./docs/WORKFLOWS.md)** - GitHub Actions explained - **[Learning System](./docs/tutorials/understanding-autonomous-workflow.md)** - How learning works --- ## Status & Monitoring The repository includes a GitHub Pages dashboard showing: - Agent activity and assignments - Workflow execution status - Learning system outputs (for review) **[View Dashboard](https://enufacas.github.io/Chained/)** --- ## Contributing Contributions are welcome. This is an experimental project, so: - **Agent definitions** - Add or improve agents in `.github/agents/` - **Workflow improvements** - Enhance orchestration patterns - **Documentation** - Help clarify how the system works - **Bug fixes** - Report and fix issues All external PRs require manual review. See [Security Implementation](./docs/SECURITY_IMPLEMENTATION.md) for details. --- ## Limitations & Known Issues - **AI-Native Control Plane**: Core system is complete and production-ready. Advanced features (multi-region, cost optimization, self-improvement) are planned enhancements. - **Mission system quality**: Generated missions often require human review for relevance and quality. The system is primarily for testing agent capabilities. - **Agent matching accuracy**: The keyword-based matching sometimes assigns inappropriate agents. Manual reassignment is possible by editing issue labels. - **A2A interactive dialogue**: Tier 1 and Tier 2 A2A are implemented (parallel execution, artifact passing). Tier 3 (real-time agent-to-agent dialogue via MCP) is in design phase. - **ADK agents**: Require GCP infrastructure (Cloud Run, Secret Manager) and API keys. See [ADK Pipeline docs](./docs/ADK_A2A_PIPELINE_IMPLEMENTATION.md). - **Gemini integration**: Requires a separate API key configuration (`GEMINI_API_KEY` secret). This is an experiment in agent orchestration patterns, not a production-ready system (except for the AI-Native Control Plane core services). --- ## License Open source. See [LICENSE](./LICENSE). --- ## Acknowledgments - **GitHub Actions** - Workflow orchestration backbone - **GitHub Copilot** - Custom agent runtime - **Gemini CLI** - Alternative agent invocation - **Model Context Protocol (MCP)** - Tool integration - **A2A Protocol** - Agent-to-agent communication standard - **Google ADK** - Agent Development Kit for cloud deployment --- *This repository is an experiment in AI agent orchestration. The patterns and conventions documented here are works in progress.*