npm i -g aiwg        # install globally
aiwg use sdlc        # deploy SDLC framework

What AIWG Is

AIWG is a cognitive architecture that gives AI coding assistants structured memory, multi-agent ensemble validation, and closed-loop self-correction. It deploys specialized agents, workflow commands, enforcement rules, and artifact templates to any of 8 AI platforms with a single CLI command.

If you have used AI coding assistants and thought "this is amazing for small tasks but falls apart on anything complex," AIWG is the missing infrastructure layer that scales AI assistance to multi-week projects.

Unlike prompt libraries or ad-hoc workflows, AIWG implements research-backed patterns from cognitive science (Miller 1956, Sweller 1988), multi-agent systems (Jacobs et al. 1991, MetaGPT, AutoGen), and software engineering (Cooper's stage-gate, FAIR Principles, W3C PROV). The system addresses the hard problems in AI-augmented development: recovering from failures, maintaining context across sessions, preventing hallucinated citations, and ensuring reproducible workflows.

What Problems Does AIWG Solve?

Base AI assistants (Claude, GPT-4, Copilot without frameworks) have three fundamental limitations:

1. No Memory Across Sessions

Each conversation starts fresh. The assistant has no idea what happened yesterday, what requirements you documented, or what decisions you made last week. You re-explain context every morning.

Without AIWG: Projects stall as context rebuilding eats time. A three-month project requires continuity, not fresh starts every session.

With AIWG: The .aiwg/ directory maintains 50-100+ interconnected artifacts across days, weeks, and months. Week 12 builds on weeks 1-11 automatically because memory persists. Agents read prior work via @-mentions instead of regenerating from scratch.

2. No Recovery Patterns

When AI generates broken code or flawed designs, you manually intervene, explain the problem, and hope the next attempt works. There is no systematic learning from failures, no structured retry, no checkpoint-and-resume.

Without AIWG: Research shows 47% of AI workflows produce inconsistent outputs without reproducibility constraints (R-LAM, Sureshkumar et al. 2026). Debugging is trial-and-error.

With AIWG: The Ralph loop implements closed-loop self-correction — execute, verify, learn from failure, adapt strategy, retry. External Ralph survives crashes and runs for 6-8+ hours autonomously. Debug memory accumulates failure patterns so the agent doesn't repeat mistakes.

3. No Quality Gates

Base assistants optimize for "sounds plausible" not "actually works." A general assistant critiques security, performance, and maintainability simultaneously — poorly. No domain specialization, no multi-perspective review, no human approval checkpoints.

Without AIWG: Production code ships without architectural review, security validation, or operational feasibility assessment.

With AIWG: 162 specialized agents provide domain expertise — Security Auditor reviews security, Test Architect reviews testability, Performance Engineer reviews scalability. Multi-agent review panels with synthesis. Human-in-the-loop gates at every phase transition. Research shows 84% cost reduction keeping humans on high-stakes decisions versus fully autonomous systems (Agent Laboratory, Schmidgall et al. 2025).

The Six Core Components

1. Memory — Structured Semantic Memory

The .aiwg/ directory is a persistent artifact repository storing requirements, architecture decisions, test strategies, risk registers, and deployment plans across sessions. This implements Retrieval-Augmented Generation patterns (Lewis et al., 2020) — agents retrieve from an evolving knowledge base rather than regenerating from scratch.

Each artifact is discoverable via @-mentions (e.g., @.aiwg/requirements/UC-001-login.md). Context sharing between agents happens through artifacts: the requirements analyst writes use cases, the architecture designer reads them.

2. Reasoning — Multi-Agent Deliberation with Synthesis

Instead of a single general-purpose assistant, AIWG provides 162 specialized agents organized by domain. Complex artifacts go through multi-agent review panels:

Architecture Document Creation:
  1. Architecture Designer drafts SAD
  2. Review Panel (3-5 agents run in parallel):
     - Security Auditor    → threat perspective
     - Performance Engineer → scalability perspective
     - Test Architect       → testability perspective
     - Technical Writer     → clarity and consistency
  3. Documentation Synthesizer merges all feedback
  4. Human approval gate → accept, iterate, or escalate

Research shows 17.9% accuracy improvement with multi-path review on complex tasks (Wang et al., GSM8K benchmarks, 2023). Agent specialization means security review is done by a security specialist, not a generalist.

3. Learning — Closed-Loop Self-Correction (Ralph)

Ralph executes tasks iteratively, learns from failures, and adapts strategy based on error patterns. Research from Roig (2025) shows recovery capability — not initial correctness — predicts agentic task success.

Ralph Iteration:
  1. Execute task with current strategy
  2. Verify results (tests pass, lint clean, types check)
  3. If failure: analyze root cause → extract structured learning → adapt strategy
  4. Log iteration state (checkpoint for resume)
  5. Repeat until success or escalate to human after 3 failed attempts

External Ralph adds crash resilience: PID file tracking, automatic restart, cross-session persistence. Tasks run for 6-8+ hours surviving terminal disconnects and system reboots.

4. Verification — Bidirectional Traceability

AIWG maintains links between documentation and code to ensure artifacts stay synchronized:

// src/auth/login.ts
/**
 * @implements @.aiwg/requirements/UC-001-login.md
 * @architecture @.aiwg/architecture/SAD.md#section-4.2
 * @tests @test/unit/auth/login.test.ts
 */
export function authenticateUser(credentials: Credentials): Promise<AuthResult> {

Verification types: Doc → Code, Code → Doc, Code → Tests, Citations → Sources. The retrieval-first citation architecture reduces citation hallucination from 56% to 0% (LitLLM benchmarks, ServiceNow 2025).

5. Planning — Phase Gates with Cognitive Load Management

AIWG structures work using Cooper's Stage-Gate methodology (1990), breaking multi-month projects into bounded phases with explicit quality criteria and human approval:

Inception → Elaboration → Construction → Transition → Production
   LOM          ABM            IOC            PR

Cognitive load optimization follows Miller's 7±2 limits (1956) and Sweller's worked examples approach (1988):

• 4 phases (not 12)
• 3-5 artifacts per phase (not 20)
• 5-7 section headings per template (not 15)
• 3-5 reviewers per panel (not 10)

6. Style — Controllable Voice Generation

Voice profiles provide continuous control over AI writing style using 12 parameters (formality, technical depth, sentence variety, jargon density, personal tone, humor, directness, examples ratio, uncertainty acknowledgment, opinion strength, transition style, authenticity markers).

Built-in voices: technical-authority (docs, RFCs), friendly-explainer (tutorials), executive-brief (summaries), casual-conversational (blogs, social). Create custom voices from your existing content with /voice-create.

A Real Project Walkthrough

Here is how the six components work together across a typical multi-week project:

Week 1: Inception

/intake-wizard "Build customer portal with real-time chat" --interactive

Memory: Intake forms capture goals, constraints, stakeholders in .aiwg/intake/
Planning: Executive Orchestrator guides through structured questionnaire
Reasoning: Requirements Analyst drafts initial use cases, Product Designer reviews UX
Verification: Requirements reference intake forms, ensuring alignment
Human Gate: Stakeholder reviews intake → approves transition to Elaboration

Weeks 2-4: Elaboration

/flow-inception-to-elaboration

Memory: Architecture doc, ADRs, threat model, test strategy accumulate in .aiwg/
Reasoning: Multi-agent review panel — Architecture Designer drafts, Security Auditor + Performance Engineer + Test Architect critique in parallel, Documentation Synthesizer merges
Learning: Ralph iterates on ADRs (generate options, evaluate against constraints, refine)
Style: Technical documents use technical-authority, stakeholder summaries use executive-brief
Human Gate: Architect reviews SAD, security team approves threat model

Weeks 5-10: Construction

/flow-elaboration-to-construction
/ralph "Implement authentication module" --completion "npm test passes"

Learning: Ralph handles implementation iterations — execute, verify (run tests), learn ("async race condition in token refresh"), adapt (add synchronization), retry
Verification: Code references requirements (@implements UC-001), tests reference code
Memory: Test plans, implementation, deployment scripts accumulate across iterations
Human Gate: Code review approves merges, QA approves test results

Weeks 11-12: Transition

/flow-deploy-to-production
/flow-hypercare-monitoring 14

Planning: Deployment checklist — monitoring, rollback plan, incident response
Learning: Ralph retries deployment steps if validation fails
Verification: Deployment scripts reference architecture (which services, what order)
Human Gate: Operations team reviews deployment plan → approves production release

Quantified Claims and Evidence

AIWG makes specific, falsifiable claims backed by peer-reviewed research:

Full references: docs/research/

When to Use AIWG (and When Not To)

Good Fit

Multi-week or multi-month projects where requirements evolve, multiple stakeholders have different concerns, quality gates are required, auditability matters, or context exceeds conversation limits.

Examples: New product features with architecture/security/operational implications, legacy system migrations requiring phased rollback strategies, research projects needing literature review and reproducibility, compliance-heavy domains (healthcare, finance, aerospace) needing audit trails.

Not the Best Fit

Single-session tasks where no memory is needed, quality gates are overkill, and overhead exceeds value.

Examples: "Write a Python script to parse this CSV," "Fix this typo," "Explain how this code works."

The Trade-off

AIWG adds structure (templates, phases, gates) that slows trivial tasks but scales to complex multi-week workflows. If your project fits in a single conversation, use a base assistant. If it spans days, weeks, or months, AIWG provides the infrastructure to maintain quality and context.

User intent → AIWG CLI → Deploy agents + rules + templates → AI platform
                │                                                │
                ▼                                                ▼
         "aiwg use sdlc"                              Claude Code / Copilot /
                │                                     Cursor / Warp / Factory /
                ▼                                     OpenCode / Codex / Windsurf
         ┌──────────────┐
         │ 188 Agents   │  Specialized AI personas with domain expertise
         │ 50 Commands  │  CLI + slash commands for workflow automation
         │ 128 Skills   │  Natural language workflow triggers
         │ 35 Rules     │  Enforcement patterns (security, quality, anti-laziness)
         │ 334 Templates│  SDLC artifact templates with progressive disclosure
         └──────────────┘
                │
                ▼
         .aiwg/ artifacts ← Persistent project memory across sessions

How It Works

AIWG orchestrates multi-agent workflows where specialized agents collaborate on complex tasks:

You: "transition to elaboration phase"

AIWG: [Step 1] Requirements Analyst   → Analyze vision document, generate use case briefs
      [Step 2] Architecture Designer  → Baseline architecture, identify technical risks     } parallel
      [Step 3] Security Architect     → Threat model, security requirements                 }
      [Step 4] Documentation Synth.   → Merge reviews into Architecture Baseline Milestone
      [Step 5] Human Gate             → GO / CONDITIONAL_GO / NO_GO decision
      [Step 6] → Next phase or iterate

The orchestration pattern: Primary Author → Parallel Reviewers → Synthesizer → Human Gate → Archive. Agents run in parallel where possible, with human-in-the-loop checkpoints at phase transitions.

Features

• 188 specialized agents — domain experts across testing, security, architecture, DevOps, cloud, frontend, backend, data engineering, documentation, and more
• 50 CLI commands — framework deployment, project scaffolding, iterative execution, metrics, reproducibility validation
• 128 workflow skills — natural language triggers for regression testing, forensics, voice profiles, quality gates, and CI/CD integration
• 35 enforcement rules — anti-laziness detection, token security, citation integrity, executable feedback, failure mitigation across 6 LLM archetypes
• 334 artifact templates — progressive disclosure templates for requirements, architecture, testing, security, deployment, and more
• 8 platform support — deploy to Claude Code, Copilot, Cursor, Warp, Factory AI, OpenCode, Codex, and Windsurf
• 6 complete frameworks — SDLC, Digital Forensics, Marketing Operations, Research Management, Media Curation, Ops Infrastructure
• 21 addons — RLM recursive decomposition, voice profiles, testing quality, mutation testing, UAT automation, and more
• Ralph Loop — iterative task execution with automatic error recovery and crash resilience (6-8 hour sessions)
• RLM addon — recursive context decomposition for processing 10M+ tokens via sub-agent delegation
• YAML metalanguage — declarative schema-validated workflow definitions (JSON Schema 2020-12)
• MCP server — Model Context Protocol integration for tool-based AI workflows
• Bidirectional traceability — @-mention system linking requirements → architecture → code → tests
• FAIR-aligned artifacts — W3C PROV provenance, GRADE quality assessment, persistent REF-XXX identifiers
• Reproducibility validation — deterministic execution modes, checkpoints, configuration snapshots

Quick Start

Prerequisites: Node.js >=18.0.0 and an AI platform (Claude Code, GitHub Copilot, Cursor, Warp Terminal, or others). See Prerequisites Guide for details.

Install & Deploy

# Install globally
npm install -g aiwg

# Deploy to your project
cd your-project
aiwg use sdlc              # Full SDLC framework (90 agents, 34 rules, 170+ templates)
aiwg use forensics         # Digital forensics & incident response (13 agents, 10 skills)
aiwg use marketing         # Marketing operations (37 agents, 87+ templates)
aiwg use media-curator     # Media archive management (6 agents, 9 commands)
aiwg use research          # Research workflow automation (8 agents, 8-stage pipeline)
aiwg use rlm               # RLM addon (recursive context decomposition)
aiwg use all               # Everything

# Or scaffold a new project
aiwg new my-project

# Check installation health
aiwg doctor

Claude Code Plugin (Alternative)

/plugin marketplace add jmagly/ai-writing-guide
/plugin install sdlc@aiwg

Multi-Platform Deployment

aiwg use sdlc                          # Claude Code (default)
aiwg use sdlc --provider copilot       # GitHub Copilot
aiwg use sdlc --provider cursor        # Cursor
aiwg use sdlc --provider warp          # Warp Terminal
aiwg use sdlc --provider factory       # Factory AI
aiwg use sdlc --provider opencode      # OpenCode
aiwg use sdlc --provider openai        # OpenAI/Codex
aiwg use sdlc --provider windsurf      # Windsurf

What You Get

Frameworks (6)

Addons (21)

Agents (188)

Specialized AI personas deployed to your platform with defined tools, responsibilities, and operating rhythms.

SDLC Agents (90)

Forensics Agents (13)

Marketing Agents (37)

Research Agents (8)

Discovery Agent, Acquisition Agent, Documentation Agent, Citation Agent, Quality Agent, Archival Agent, Provenance Agent, Workflow Agent

Media Curator Agents (6)

Discography Analyst, Source Discoverer, Acquisition Manager, Quality Assessor, Metadata Curator, Completeness Tracker

Rules (35)

Enforcement patterns that prevent common AI failure modes. Rules deploy automatically with their framework.

Core Rules (10) — Always Active

SDLC Rules (34) — Active with Framework

Actionable feedback, mention wiring, HITL gates, agent fallback, provenance tracking, TAO loop, reproducibility validation, SDLC orchestration, agent-friendly code, agent generation guardrails, artifact discovery, HITL patterns, human gate display, thought protocol, reasoning sections, few-shot examples, best output selection, reproducibility, progressive disclosure, conversable agent interface, auto-reply chains, criticality panel sizing, qualified references.

Research Rules (2) — Active with Research

Research metadata (FAIR-compliant YAML frontmatter), index generation (auto-generated INDEX.md per FAIR F4).

Skills (128)

Natural language workflow triggers. Say "what's the project status?" and the project-awareness skill activates.

Framework Deep Dives

SDLC Complete — Full Software Development Lifecycle

The SDLC framework implements a phase-gated development lifecycle with 90 specialized agents, 34 enforcement rules, and 170+ artifact templates. Natural language commands drive phase transitions with automated quality gates.

 ┌──────────┐    ┌─────────────┐    ┌──────────────┐    ┌────────────┐    ┌────────────┐
 │ CONCEPT  │───▶│  INCEPTION  │───▶│ ELABORATION  │───▶│CONSTRUCTION│───▶│ TRANSITION │
 │          │    │             │    │              │    │            │    │            │
 │ Intake   │    │ Vision      │    │ Architecture │    │ Code       │    │ Deploy     │
 │ Wizard   │    │ Requirements│    │ Risk Retire  │    │ Test       │    │ Hypercare  │
 │ Solution │    │ Stakeholder │    │ Prototype    │    │ Review     │    │ Handoff    │
 │ Profile  │    │ Analysis    │    │ API Design   │    │ Iterate    │    │ Knowledge  │
 └──────────┘    └──────┬──────┘    └──────┬───────┘    └─────┬──────┘    └────────────┘
                        │                  │                   │
                     ┌──▼──┐            ┌──▼──┐            ┌──▼──┐
                     │ LOM │            │ ABM │            │ IOC │    ← Quality Gates
                     │Gate │            │Gate │            │Gate │
                     └─────┘            └─────┘            └─────┘

 LOM = Lifecycle Objectives Milestone    ABM = Architecture Baseline Milestone
 IOC = Initial Operational Capability

SDLC Flow Commands (24):

SDLC Accelerate — Idea to Construction-Ready in One Command:

# From a description
aiwg sdlc-accelerate "AI-powered code review tool with GitHub integration"

# From existing codebase
aiwg sdlc-accelerate --from-codebase .

# Resume interrupted pipeline
aiwg sdlc-accelerate --resume

Generates intake form, vision document, use cases, architecture baseline, risk register, test strategy, and deployment plan — all with human approval gates between phases.

Dual-Track Iteration Model:

        ┌─────────────────────────────────────────────────┐
        │                ITERATION N                       │
        │                                                  │
        │  Discovery Track          Delivery Track         │
        │  (Next iteration)         (Current iteration)    │
        │                                                  │
        │  ┌─────────────┐          ┌──────────────┐      │
        │  │ Requirements│          │ Implement    │      │
        │  │ Research    │          │ Test         │      │
        │  │ Design      │          │ Review       │      │
        │  │ Validate    │          │ Deploy       │      │
        │  └─────┬───────┘          └──────┬───────┘      │
        │        │                         │               │
        │        └────────────┬────────────┘               │
        │                     │                            │
        │              ┌──────▼──────┐                     │
        │              │ Iteration   │                     │
        │              │ Assessment  │                     │
        │              └─────────────┘                     │
        └─────────────────────────────────────────────────┘

Metrics & Quality Tracking:

Forensics Complete — Digital Forensics & Incident Response

Full DFIR investigation workflow following NIST SP 800-86, with MITRE ATT&CK mapping and Sigma rule hunting.

 ┌──────────┐    ┌──────────┐    ┌────────────┐    ┌──────────┐    ┌──────────┐
 │  SCOPE   │───▶│  TRIAGE  │───▶│  ACQUIRE   │───▶│ ANALYZE  │───▶│  REPORT  │
 │          │    │          │    │            │    │          │    │          │
 │ Profile  │    │ Volatile │    │ Evidence   │    │ Log      │    │ Executive│
 │ target   │    │ data     │    │ collection │    │ Timeline │    │ summary  │
 │ system   │    │ capture  │    │ Chain of   │    │ IOC      │    │ Findings │
 │          │    │ RFC 3227 │    │ custody    │    │ Sigma    │    │ Timeline │
 └──────────┘    └──────────┘    └────────────┘    └──────────┘    └──────────┘
                                       │
                                  SHA-256 hash
                                  verification

Investigation Commands:

/forensics-profile          # Build target system profile via SSH
/forensics-triage           # Quick triage following RFC 3227 volatility order
/forensics-acquire          # Evidence acquisition with chain of custody
/forensics-investigate      # Full multi-agent investigation workflow
/forensics-timeline         # Build correlated event timeline
/forensics-hunt             # Threat hunt using Sigma rules
/forensics-ioc              # Extract and enrich IOCs
/forensics-report           # Generate forensic investigation report
/forensics-status           # Show investigation dashboard

Bundled Sigma Rules (8):

Supported Evidence Sources:

Media/Marketing Kit — Campaign Lifecycle

 ┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────────┐
 │ STRATEGY │───▶│ CREATION │───▶│  REVIEW  │───▶│ PUBLISH  │───▶│ ANALYZE  │
 │          │    │          │    │          │    │          │    │          │
 │ Research │    │ Copy     │    │ Brand    │    │ Schedule │    │ KPIs     │
 │ Audience │    │ Design   │    │ Legal    │    │ Channels │    │ Reports  │
 │ Strategy │    │ Content  │    │ Quality  │    │ Launch   │    │ ROI      │
 └──────────┘    └──────────┘    └──────────┘    └──────────┘    └──────────┘

37 agents across strategy, creation, management, analytics, and communications. 87+ templates covering campaign intake, brand guidelines, content briefs, social playbooks, email sequences, PR kits, and analytics dashboards.

Media Curator — Intelligent Archive Management

# Full curation pipeline
/curate "Pink Floyd"

# Step by step
/analyze-artist "Pink Floyd"           # Identify eras, catalog structure
/find-sources "Pink Floyd" "DSOTM"     # Discover across YouTube, Archive.org, Bandcamp
/acquire                               # Download with format selection
/tag-collection                        # Apply metadata, embed artwork, rename
/check-completeness                    # Gap analysis against canonical discography
/assemble "Pink Floyd live 1973"       # Build thematic compilations
/export --format plex                  # Export to Plex, Jellyfin, MPD, or archival
/verify-archive                        # SHA-256 integrity verification

Quality tiers: Tier 1 (Official/Lossless) → Tier 2 (High Quality) → Tier 3 (Acceptable) → Tier 4 (Avoid). Standards: ID3v2.4, Vorbis Comments, MusicBrainz, PREMIS 3.0, W3C PROV-O.

Research Complete — Academic Research Pipeline

 ┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────────┐
 │ DISCOVER │───▶│ ACQUIRE  │───▶│ DOCUMENT │───▶│ ARCHIVE  │
 │          │    │          │    │          │    │          │
 │ Search   │    │ Download │    │ RAG      │    │ OAIS     │
 │ databases│    │ PDF      │    │ summaries│    │ lifecycle│
 │ Rank     │    │ Metadata │    │ Citations│    │ FAIR     │
 │ results  │    │ extract  │    │ GRADE    │    │ W3C PROV │
 └──────────┘    └──────────┘    └──────────┘    └──────────┘

8-stage pipeline: Discovery → Acquisition → Documentation → Citation → Quality Assessment → Synthesis → Gap Analysis → Archival. Persistent REF-XXX identifiers. GRADE scoring (HIGH/MODERATE/LOW/VERY LOW). Unpaywall integration for open access papers.

Voice Framework — Content Voice Consistency

4 built-in voice profiles with create, analyze, blend, and apply skills:

# Apply a voice to content
"Apply technical-authority voice to docs/architecture.md"

# Create a custom voice from existing content
/voice-create --source "blog-posts/*.md" --name "company-voice"

# Analyze writing patterns in a voice sample
/voice-analyze docs/existing-content.md

# Blend two voices
/voice-blend technical-authority casual-conversational --ratio 70:30

# Detect AI patterns and suggest authentic alternatives
/ai-pattern-detection docs/generated-content.md

MCP Server — Model Context Protocol Integration

AIWG includes a built-in MCP server for tool-based AI workflow integration:

# Start MCP server
aiwg mcp serve

# Install into Claude Desktop
aiwg mcp install claude

# Show capabilities
aiwg mcp info

The MCP server exposes AIWG's artifact management, workflow execution, and project health capabilities as tools that any MCP-compatible AI platform can invoke programmatically.

Agent Evaluation Framework

Test agent quality with archetype resistance testing based on Roig (2025) failure patterns:

# Evaluate a specific agent
/eval-agent security-auditor

# Test archetype resistance
/eval-agent test-engineer --category archetype

# Run performance benchmarks
/eval-agent code-reviewer --category performance

Test Categories:

Target score: >=85% per agent. Results include passed/failed breakdown with evidence.

Bidirectional Traceability — @-Mention System

Link requirements to architecture to code to tests with semantic @-mentions:

<!-- In a use case document -->
This use case @implements(UC-001) the authentication flow
described in @architecture(SAD-section-3.2).

<!-- In a test file -->
// @tests(UC-001) @depends(auth-service)
describe('authentication flow', () => { ... });

Mention Commands:

/mention-wire         # Analyze codebase and inject @-mentions for traceability
/mention-validate     # Validate all @-mentions resolve to existing files
/mention-report       # Generate traceability report
/mention-lint         # Lint @-mentions for style consistency
/mention-conventions  # Display naming conventions and placement rules

Relationship qualifiers: @implements, @tests, @depends, @derives-from, @blocked-by, @supersedes. Enables queries like "what implements UC-001?" and "what tests cover the auth module?"

Configuration & Customization

Workspace Structure

your-project/
├── .aiwg/                    # SDLC artifacts (persistent project memory)
│   ├── intake/               # Project intake forms
│   ├── requirements/         # Use cases, user stories, NFRs
│   ├── architecture/         # SAD, ADRs, system diagrams
│   ├── planning/             # Phase plans, iteration plans
│   ├── risks/                # Risk register, mitigations
│   ├── testing/              # Test strategy, plans, results
│   ├── security/             # Threat models, security gates
│   ├── deployment/           # Deployment plans, runbooks
│   ├── reports/              # Generated status reports
│   ├── ralph/                # In-session Ralph loop state
│   ├── ralph-external/       # External Ralph crash-resilient state
│   ├── research/             # Research corpus and findings
│   ├── forensics/            # Investigation artifacts
│   ├── working/              # Temporary files (safe to delete)
│   └── frameworks/           # Installed framework registry
│       └── registry.json
├── .claude/                  # Claude Code deployment
│   ├── agents/               # 162 agent definitions
│   ├── commands/             # Slash commands
│   ├── skills/               # 86 skill definitions
│   └── rules/                # RULES-INDEX.md + on-demand full rules
├── .github/                  # GitHub Copilot deployment
├── .cursor/                  # Cursor deployment
├── .warp/                    # Warp Terminal deployment
└── CLAUDE.md                 # Project instructions (auto-generated)

Creating Custom Extensions

# Add a custom agent
aiwg add-agent my-domain-expert

# Add a custom command
aiwg add-command my-workflow

# Add a custom skill
aiwg add-skill my-capability

# Scaffold a complete addon
aiwg scaffold-addon my-addon

# Scaffold a complete framework
aiwg scaffold-framework my-framework

# Validate all extension metadata
aiwg validate-metadata

Artifact Index & Discovery

# Build searchable artifact index
aiwg index build
aiwg index build --force --verbose

# Search artifacts by keyword
aiwg index query "authentication" --json

# Show dependency graph for an artifact
aiwg index deps .aiwg/requirements/UC-001.md --json

# Index statistics
aiwg index stats --json

The index supports multiple graphs: project graph (.aiwg/ artifacts) and framework graph (agentic/code/ + docs/).

Doc Sync — Bidirectional Documentation

# Audit doc drift (dry run)
aiwg doc-sync code-to-docs --dry-run

# Sync docs to match code
aiwg doc-sync code-to-docs

# Bidirectional reconciliation
aiwg doc-sync full --interactive

Reproducibility Validation

# Show/set execution mode (strict = temperature 0, fixed seed)
aiwg execution-mode

# Create execution snapshot
aiwg snapshot

# Create workflow checkpoint
aiwg checkpoint

# Validate workflow reproducibility
aiwg reproducibility-validate

Thresholds: compliance audit (100%), security scan (100%), test generation (95%).

Issue-Driven Development

AIWG integrates with issue trackers for 2-way human-AI collaboration:

# Create issues from any backend (Gitea, GitHub, Jira, Linear, local files)
/issue-create "Implement OAuth2 flow" --labels "feature,auth"

# List and filter issues
/issue-list --state open --labels "priority:high"

# Drive an issue with Ralph loop — posts status to issue thread
/issue-driven-ralph 42

# Auto-sync issues from commits and artifacts
/issue-sync

# Close with comprehensive summary and verification
/issue-close 42

The /address-issues command orchestrates issue-thread-driven Ralph loops with automatic progress posting and human feedback incorporation at each cycle.

Daemon Mode & Messaging Integration

Daemon Mode

# Background file watching, cron scheduling, IPC
aiwg daemon start

See Daemon Guide for background agent orchestration.

Messaging Integration

Bidirectional Slack, Discord, and Telegram bots for remote agent control:

# Connect to messaging platforms
aiwg messaging connect slack
aiwg messaging connect discord
aiwg messaging connect telegram

See Messaging Guide for setup and configuration.

See It In Action

# Generate project intake from natural language
/intake-wizard "Build customer portal with real-time chat"

# Accelerate from idea to construction-ready
/sdlc-accelerate "AI-powered code review tool"

# Phase transition with automated gate check
/flow-inception-to-elaboration

# Iterative task execution — "iteration beats perfection"
/ralph "Fix all failing tests" --completion "npm test passes"

# Long-running tasks with crash recovery (6-8 hours)
/ralph-external "Migrate to TypeScript" --completion "npx tsc --noEmit exits 0"

# Process massive codebases with recursive context decomposition
/rlm-query "src/**/*.ts" "Extract all exported interfaces" --model haiku
/rlm-batch "src/components/*.tsx" "Add TypeScript types" --parallel 4

# Digital forensics investigation
/forensics-investigate
/forensics-triage
/forensics-timeline

# Scan codebase for agent-readiness
/codebase-health --format text

# Decompose large files into agent-friendly modules
/decompose-file src/large-file.ts --execute

# Deploy to production with rollback gates
/flow-deploy-to-production

# Security assessment
/security-audit

# Voice transformation
"Apply technical-authority voice to docs/architecture.md"
"Create a voice profile based on our existing blog posts"

Platform Support

All 8 platforms receive agents, commands, skills, and rules. Deployment adapts to each platform's conventions automatically.

CLI Reference (50 Commands)

Quick Reference

# Deploy frameworks
aiwg use sdlc                    # SDLC framework
aiwg use forensics               # Forensics framework
aiwg use all                     # Everything
aiwg use sdlc --provider copilot # Deploy to GitHub Copilot

# Project management
aiwg new my-project              # Scaffold new project
aiwg status                      # Workspace health
aiwg doctor                      # Installation diagnostics

# Iterative execution (Ralph Loop)
aiwg ralph "Fix all tests" --completion "npm test passes"
aiwg ralph-status                # Check loop progress
aiwg ralph-abort                 # Cancel running loop
aiwg ralph-resume                # Resume interrupted loop
aiwg ralph-external "Migrate to TS" --completion "tsc --noEmit exits 0"

# Artifact discovery
aiwg index build                 # Build artifact index
aiwg index query "authentication" --json
aiwg index deps .aiwg/requirements/UC-001.md --json

# Documentation sync
aiwg doc-sync code-to-docs --dry-run
aiwg doc-sync full --interactive

# Metrics
aiwg cost-report                 # Session cost breakdown
aiwg metrics-tokens              # Token usage

# SDLC accelerate
aiwg sdlc-accelerate "Project description"
aiwg sdlc-accelerate --from-codebase .

Architecture

Extension System

AIWG uses a unified extension system with 10 extension types, all deployable across 8 platforms:

Multi-Agent Orchestration

                    ┌─────────────────────┐
                    │ Executive Orchestrator│
                    └──────────┬──────────┘
                               │
              ┌────────────────┼────────────────┐
              ▼                ▼                 ▼
      ┌──────────────┐ ┌──────────────┐ ┌──────────────┐
      │Primary Author│ │  Reviewer 1  │ │  Reviewer 2  │  ← Parallel
      │(e.g. Req.   │ │(e.g. Security│ │(e.g. Test    │
      │  Analyst)    │ │  Architect)  │ │  Architect)  │
      └──────┬───────┘ └──────┬───────┘ └──────┬───────┘
             └────────────────┼────────────────┘
                              ▼
                    ┌─────────────────────┐
                    │  Documentation      │
                    │  Synthesizer        │  ← Merge all reviews
                    └──────────┬──────────┘
                               ▼
                    ┌─────────────────────┐
                    │  Human Gate         │  ← GO / NO_GO decision
                    └──────────┬──────────┘
                               ▼
                    ┌─────────────────────┐
                    │  .aiwg/ Archive     │  ← Persistent artifacts
                    └─────────────────────┘

YAML Metalanguage

AIWG is pioneering a declarative YAML metalanguage for multi-agent workflow orchestration. Schema-validated YAML defines agent topology, workflow DAGs, gate conditions, and artifact contracts — while natural language handles behavioral logic.

# Example: flow definition (schema-validated)
flow:
  id: inception-to-elaboration
  model: opus
  entry_criteria:
    gate: LOM
    artifacts:
      - path: .aiwg/requirements/vision-document.md
        required: true
  steps:
    - id: requirements-analysis
      agent: requirements-analyst
      parallel_group: reviews
    - id: architecture-baseline
      agent: architecture-designer
      parallel_group: reviews
    - id: synthesis
      agent: documentation-synthesizer
      depends_on: [requirements-analysis, architecture-baseline]
  exit_criteria:
    gate: ABM
    decision: [GO, CONDITIONAL_GO, NO_GO]

JSON Schema definitions for flow.yaml, agent.yaml, rule.yaml, and skill.yaml at agentic/code/frameworks/sdlc-complete/schemas/metalanguage/.

Project Artifacts (.aiwg/)

All SDLC artifacts persist in .aiwg/ — structured project memory that survives across AI sessions:

.aiwg/
├── intake/           # Project intake forms, solution profiles
├── requirements/     # Use cases, user stories, NFRs
├── architecture/     # SAD, ADRs, diagrams
├── planning/         # Phase plans, iteration plans
├── risks/            # Risk register, mitigations
├── testing/          # Test strategy, test plans
├── security/         # Threat models, security gates
├── deployment/       # Deployment plans, runbooks
├── reports/          # Generated status reports
├── ralph/            # Ralph loop state and history
└── frameworks/       # Installed framework registry

Ralph Loop — Autonomous Long-Running Agent Orchestration

The Ralph Loop is the core execution philosophy: iteration beats perfection. Instead of getting everything right on the first attempt, the agent executes in a retry loop where errors become learning data. Ralph supports both in-session loops and crash-resilient external loops that run indefinitely — surviving process crashes, terminal disconnects, and system reboots.

In-Session Ralph (Minutes to Hours)

# Iterative task execution with automatic error recovery
/ralph "Fix all failing tests" --completion "npm test passes with 0 failures"
/ralph "Reach 80% coverage" --completion "coverage report shows >80%" --max-iterations 20

# Issue-driven Ralph — posts cycle status to issue threads, incorporates human feedback
/issue-driven-ralph 42    # Drives issue #42 with 2-way human-AI collaboration

External Ralph — Crash-Resilient Autonomous Agents (Hours to Days)

External Ralph runs as a persistent background process with PID file tracking, crash recovery, and automatic restart. The agent continues working even if your terminal disconnects or the host reboots.

# Long-running autonomous task (6-8+ hours, survives crashes)
/ralph-external "Migrate entire codebase to TypeScript" \
  --completion "npx tsc --noEmit exits 0" \
  --timeout 480

# Autonomous code review loop
/ralph-external "Review and fix all security vulnerabilities" \
  --completion "npm audit shows 0 vulnerabilities"

# Continuous integration loop
/ralph-external "Get all tests passing on Node 18 and 22" \
  --completion "npm test passes on both versions"

External Ralph features:

• Crash resilience — PID file recovery, automatic restart on process death
• Checkpoint system — saves progress at each iteration boundary, resumes from last checkpoint
• Cross-session persistence — state stored in .aiwg/ralph-external/, survives terminal disconnects
• Debug memory — learns from failure patterns across iterations, applies lessons to subsequent attempts
• Episodic memory — /ralph-reflect shows accumulated learnings and strategy evolution
• Completion reports — detailed iteration history saved to .aiwg/ralph/

Scheduled and Remote Agents

# Schedule recurring autonomous agent tasks
/schedule create "Run security audit" --cron "0 9 * * 1"    # Every Monday 9am
/schedule create "Check dependency updates" --cron "0 0 * *" # Monthly

# Remote agent triggers — execute on schedule from anywhere
/schedule list
/schedule run <trigger-id>

Ralph Control

/ralph-status     # Check current/previous loop status
/ralph-resume     # Resume interrupted loop from last checkpoint
/ralph-abort      # Cancel running loop (optionally revert changes)
/ralph-memory     # View debug memory entries and failure patterns
/ralph-reflect    # View episodic memory and strategy evolution
/ralph-analytics  # Execution metrics and performance history

How It Works

Each iteration follows the TAO loop (Thought → Action → Observation):

Iteration N:
  1. THINK  — Analyze current state + accumulated learnings from iterations 1..N-1
  2. ACT    — Make changes based on task + debug memory + failure patterns
  3. VERIFY — Run completion command (tests, build, lint, coverage, etc.)
  4. LEARN  — If verification fails, extract root cause → store in debug memory
  5. DECIDE — Pass? → Complete. Fail? → Iterate. Max retries? → Escalate to human.

The debug memory system implements executable feedback: the agent doesn't just retry — it learns what went wrong and why, then applies that knowledge to the next attempt. After 3 failed attempts at the same root cause, it escalates to a human rather than looping forever.

Research foundation: Self-Refine (Madaan et al., NeurIPS 2023), ReAct (Yao et al., ICLR 2023), METR 2025 (recovery capability dominates agentic task success), Reflexion (Shinn et al., 2023).

RLM — Recursive Context Decomposition

Process codebases and documents far beyond any model's context window:

# Query: fan-out across files, gather results
/rlm-query "src/**/*.ts" "Extract all exported interfaces" --model haiku

# Batch: parallel processing with configurable concurrency
/rlm-batch "src/components/*.tsx" "Add TypeScript types" --parallel 4

# Status: monitor decomposition progress
/rlm-status

The RLM addon decomposes large inputs into chunks, delegates each to a sub-agent, and synthesizes results. Processes 10M+ tokens through recursive delegation.

Research foundation: Recursive Language Models (Zhang, Kraska, Khattab — MIT CSAIL, 2026).

Research Foundations

AIWG's architecture is grounded in peer-reviewed research across cognitive science, multi-agent systems, software engineering, and AI safety. Full corpus: research-papers (168 papers). Entries ordered highest to lowest GRADE evidence quality within each category.

Cognitive Foundations

• Miller, G.A. (1956). The Magical Number Seven, Plus or Minus Two. Psychological Review, 63(2), 81–97. doi:10.1037/h0043158
• Sweller, J. (1988). Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science, 12(2), 257–285. doi:10.1207/s15516709cog1202_4
• Anderson, J.R. et al. (2004). An Integrated Theory of the Mind. Psychological Review, 111(4), 1036–1060. (ACT-R cognitive architecture)
• Laird, J.E., Newell, A. & Rosenbloom, P.S. (1987). SOAR: An Architecture for General Intelligence. Artificial Intelligence, 33(1), 1–64.
• Harel, D. (1987). Statecharts: A Visual Formalism for Complex Systems. Science of Computer Programming, 8(3), 231–274.
• Young, S. et al. (2010). The Hidden Information State Model: A Practical Framework for POMDP-Based Spoken Dialogue Management. Computer Speech & Language, 24(2), 150–174.

Multi-Agent Systems & Orchestration

• Jacobs, R.A. et al. (1991). Adaptive Mixtures of Local Experts. Neural Computation, 3(1), 79–87. (Mixture-of-Experts foundation)
• Hong, S. et al. (2024). MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework. ICLR 2024. (85.9% HumanEval, SOP-based orchestration)
• Qian, C. et al. (2024). ChatDev: Communicative Agents for Software Development. ACL 2024.
• Shen, Y. et al. (2023). HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in HuggingFace. NeurIPS 2023.
• Tao, W. et al. (2024). MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution.
• Zhang, J. et al. (2025). AFlow: Automating Agentic Workflow Generation. ICLR 2025 Oral. (5.7% avg gain over best manual methods)
• Wu, Q. et al. (2023). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. (Conversational multi-agent framework)
• Yu, C. et al. (2025). A Survey on Agent Workflow — Status and Future. (24 systems, 11 metrics)
• Lodha, D. et al. (2026). MCP-Diag: A Deterministic, Protocol-Driven Architecture for AI-Native Network Diagnostics. COMSNETS 2026. (First production MCP system)
• Yu, G. (2026). AdaptOrch: Adaptive Orchestration for Multi-Agent LLM Systems Through Topology-Aware Task Planning. (12–23% improvement across 4 topologies)
• Gerred (2025). Multi-Agent Orchestration. Tool isolation, resource boundaries, observable coordination.
• Falconer, S. (2025). Event-Driven Multi-Agent Systems. Confluent. 4 Kafka orchestration patterns.
• Mario, M. (2025). Multi-Agent System Patterns: A Unified Guide to Designing Agentic Architectures. 4-dimensional framework.
• Runkle, S. (2026). Choosing the Right Multi-Agent Architecture. LangChain. Subagents, skills, handoffs, 90.2% improvement stat.
• Towards Data Science (2025). Why Your Multi-Agent System Is Failing: Escaping the 17x Error Trap. 17.2x error amplification, 4-agent coordination threshold.
• NexAI Tech (2025). Multi-AI Agent Architecture Patterns for Scale. Enterprise 5-layer architecture, 3 orchestration patterns.
• Wexford, E. (2026). How to Build Multi-Agent Systems: Complete 2026 Guide. DEV Community. 3–7 agents optimal sizing.

Reasoning & Planning

• Wei, J. et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022.
• Wang, X. et al. (2023). Self-Consistency Improves Chain of Thought Reasoning in Language Models. ICLR 2023.
• Yao, S. et al. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.
• Yao, S. et al. (2023). Tree of Thoughts: Deliberate Problem Solving with Large Language Models. NeurIPS 2023.
• Zhou, A. et al. (2024). Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models. ICML 2024.
• Kojima, T. et al. (2022). Large Language Models are Zero-Shot Reasoners. NeurIPS 2022. ("Let's think step by step")
• Liu, Z. et al. (2026). Exploratory Memory-Augmented LLM Agent via Hybrid On- and Off-Policy Optimization (EMPO²). ICLR 2026. (128.6% over GRPO on ScienceWorld)

Self-Correction & Iterative Refinement

• Madaan, A. et al. (2023). Self-Refine: Iterative Refinement with Self-Feedback. NeurIPS 2023. (+4.2% HumanEval, −63% revision cost)
• Shinn, N. et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS 2023.

Stage-Gate, SDLC & Traceability

• Cooper, R.G. (1990). Stage-Gate Systems: A New Tool for Managing New Products. Business Horizons, 33(3), 44–54.
• Jacobson, I., Booch, G. & Rumbaugh, J. (1999). The Unified Software Development Process. Addison-Wesley. ISBN 978-0-201-57169-1.
• Gotel, O.C.Z. & Finkelstein, A.C.W. (1994). An Analysis of the Requirements Traceability Problem. IEEE ICRE 1994.

Software Engineering & Agent-Computer Interface

• Jimenez, C.E. et al. (2024). SWE-bench: Can Language Models Resolve Real-world GitHub Issues?. ICLR 2024.
• Wang, X. et al. (2024). Executable Code Actions Elicit Better LLM Agents (CodeAct). ICML 2024. (Up to 20% higher success rate)
• Yang, J. et al. (2024). SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering. NeurIPS 2024. (12.47% SWE-bench)
• Laurent, A. (2025). A Comparison of AI Code Assistants for Large Codebases. IntuitionLabs. (62% AI code contains flaws)
• Augment Code (2025). AI Coding Assistants for Large Codebases: A Complete Guide. (73% compile locally but violate patterns)
• AlgoMaster (2025). How to Use AI Effectively in Large Codebases. Retrieval as bottleneck framing.

Context Engineering & Memory

• Liu, N.F. et al. (2024). Lost in the Middle: How Language Models Use Long Contexts. TACL 12, 157–173. doi:10.1162/tacl_a_00638
• Dai, Y. et al. (2025). Pretraining Context Compressor for Large Language Models with Embedding-Based Memory. ACL 2025.
• Kang, M. et al. (2025). ACON: Optimizing Context Compression for Long-Horizon LLM Agents. (26–54% peak token reduction, >95% accuracy preserved)
• Liu, F. & Qiu, H. (2025). Context Cascade Compression (C3): Exploring the Upper Limits of Text Compression. (98% precision at 20x compression)
• Vasilopoulos, A. (2026). Codified Context: Infrastructure for AI Agents in a Complex Codebase. (Three-tier context infrastructure: constitution + 19 agents + 34-doc KB)
• Ostby, D.L. (2025). Stingy Context: Compressing Code Context for Cost-Effective AI Development Assistance. (TREEFRAG, 18:1 compression ratio)
• Anthropic Applied AI Team (2026). Effective Context Engineering for AI Agents. Anthropic Engineering Blog.
• Huang, J.Y. et al. (2026). Do LLMs Benefit From Their Own Words? (36.4% of multi-turn prompts self-contained; up to 10x context reduction)
• Böckeler, B. (2026). Context Engineering for Coding Agents. Martin Fowler's Blog. Two-category framework.
• Haseeb, M. (2025). Context Engineering for Multi-Agent LLM Code Assistants. (80% vs 40% single-shot success)
• Verma, N. (2026). Focus Agent: LLM Agent with Active Context Compression for SWE-Bench. (22.7% token reduction via consolidate/withdraw)
• Zylos Research (2026). Long-Running AI Agents and Task Decomposition. (35-min degradation threshold, Planner-Worker model)
• Zylos Research (2026). LLM Context Window Management and Long-Context Strategies. (Lost-in-Middle persists, TTT-E2E 35× speedup)

Agent Memory & Knowledge Systems

• Laird, J.E. et al. (1987). SOAR: An Architecture for General Intelligence. Artificial Intelligence, 33(1), 1–64.
• Anderson, J.R. et al. (2004). An Integrated Theory of the Mind (ACT-R). Psychological Review, 111(4), 1036–1060.
• Park, J.S. et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023. doi:10.1145/3586183.3606763
• Xu, W. et al. (2025). A-MEM: Agentic Memory for LLM Agents. (Zettelkasten-inspired, 85–93% token reduction)
• Hu, Y. et al. (2025). Memory in the Age of AI Agents: A Survey. (Surveys 100+ implementations, forms-functions-dynamics framework)
• Rezazadeh, A. et al. (2025). Collaborative Memory: Multi-User Memory Sharing in LLM Agents with Dynamic Access Control. (61% resource reduction)
• Yuen, S. et al. (2025). Intrinsic Memory Agents: Heterogeneous Multi-Agent LLM Systems through Structured Contextual Memory. Role-aligned heterogeneous memory.
• Graves, A., Wayne, G. & Danihelka, I. (2014). Neural Turing Machines. External memory architectures.
• Packer, C. et al. (2023). MemGPT: Towards LLMs as Operating Systems. OS-inspired virtual context paging.
• Yu, Z. et al. (2026). Multi-Agent Memory from a Computer Architecture Perspective. Architecture 2.0 '26. Three-layer I/O-cache-memory hierarchy.
• Chhikara, P. et al. (2025). Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory. (26% accuracy gain, 91% latency reduction)

Recursive Context Decomposition

• Zhang, A.L., Kraska, T. & Khattab, O. (2026). Recursive Language Models. arXiv:2512.24601. MIT CSAIL. (10M+ token processing, up to 3x cheaper than summarization)

Provenance, Reproducibility & Research Management

• Wilkinson, M.D. et al. (2016). The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data, 3, 160018. (G20, EU, NIH endorsement)
• W3C (2013). PROV-DM: The PROV Data Model. W3C Recommendation.
• CCSDS (2024). Reference Model for an Open Archival Information System (OAIS). ISO 14721. (Digital preservation lifecycle)
• GRADE Working Group (2004–present). GRADE Handbook. Evidence quality assessment. Adopted by WHO, Cochrane, NICE, and 100+ organizations.
• Schmidgall, S. et al. (2025). Agent Laboratory: Using LLM Agents as Research Assistants. (84% cost reduction)
• Sureshkumar, V. et al. (2026). R-LAM: Towards Reproducibility in Large Action Model Workflows. (47% of workflows non-reproducible without constraints)
• ServiceNow Research (2025). LitLLM for Scientific Literature Reviews. RAG-based literature review, no hallucination approach.

AI Safety & Failure Modes

• Tang, R. et al. (2023). Large Language Models Can be Lazy Learners: Analyze Shortcuts in In-Context Learning. ACL 2023 Findings. doi:10.18653/v1/2023.findings-acl.284
• Bandara, E. et al. (2025). A Practical Guide for Designing, Developing, and Deploying Production-Grade Agentic AI Workflows.
• Roig, J.V. (2025). How Do LLMs Fail In Agentic Scenarios? A Qualitative Analysis.
• Von Arx, S., Chan, L. & Barnes, E. (2025). Recent Frontier Models Are Reward Hacking. METR Research Blog.
• Anthropic Alignment Team (2025). From shortcuts to sabotage: Natural emergent misalignment from reward hacking. Anthropic Research.
• Batista, R.M. & Griffiths, T.L. (2026). A Rational Analysis of the Effects of Sycophantic AI. (Bayesian analysis, epistemic risk)
• Kumar, R.S.S. et al. (2025). Taxonomy of Failure Modes in Agentic AI Systems. Microsoft Security.
• Twiss, J. (2026). AI Coding Degrades: Silent Failures Emerge. IEEE Spectrum.
• van Linschoten, A.S. (2025). The Agent Deployment Gap: Why Your LLM Loop Isn't Production-Ready. ZenML.

Task Decomposition & Declarative Pipelines

• Khot, T. et al. (2023). Decomposed Prompting: A Modular Approach for Solving Complex Tasks. ICLR 2023.
• Khattab, O. et al. (2023). DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines. ICLR 2024.
• Dohan, D. et al. (2022). Language Model Cascades. Google Brain. (PGM formalization of multi-step LLM pipelines)
• Peng, B. et al. (2023). Check Your Facts and Try Again: Improving Large Language Models with External Knowledge and Automated Feedback. Microsoft Research.

Training, Alignment & In-Context Learning

• Ouyang, L. et al. (2022). Training language models to follow instructions with human feedback (InstructGPT). (RLHF methodology)
• Dong, Q. et al. (2024). A Survey on In-Context Learning. EMNLP 2024.
• Bai, Y. et al. (2022). Constitutional AI: Harmlessness from AI Feedback. (Principle-based alignment)

RAG, Retrieval & Tool Use

• Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020.
• Schick, T. et al. (2023). Toolformer: Language Models Can Teach Themselves to Use Tools. (Self-supervised tool learning)

Domain Knowledge & Specialization

• Song, Z. et al. (2025). Injecting Domain-Specific Knowledge into Large Language Models: A Comprehensive Survey. EMNLP 2025 Findings. (Four-paradigm taxonomy)
• Zhang, T. et al. (2024). RAFT: Adapting Language Model to Domain Specific RAG. COLM 2024. (Outperforms RAG-only and SFT-only)

Constrained Generation & Output Validation

• Beurer-Kellner, L., Fischer, M. & Vechev, M. (2023). Prompting Is Programming: A Query Language for Large Language Models (LMQL). PLDI 2023. doi:10.1145/3591300 (26–85% token reduction)
• Willard, B.T. & Louf, R. (2023). Efficient Guided Generation for Large Language Models (Outlines). (0% parse failures by construction)
• Lhoest, Q. & Turuta, M. (2024). Structured Generation with Outlines. Hugging Face Blog. (1.5–3x speedup)
• Gerganov, G. et al. (2024). Grammar-Based Sampling (GBNF) — llama.cpp. Context-free grammar constrained sampling.

LLM Serving & Local Deployment

• Yu, G. et al. (2022). Orca: A Distributed Serving System for Transformer-Based Generative Models. OSDI '22. (36.9x throughput improvement)
• Kwon, W. et al. (2023). Efficient Memory Management for Large Language Model Serving with PagedAttention. SOSP '23. UC Berkeley. (2–4x throughput vs HuggingFace)
• Ollama Team (2024). Ollama Concurrent Requests and Performance FAQ. OLLAMA_NUM_PARALLEL configuration guidance.

MCP & Agentic Standards

• Agentic AI Foundation / Linux Foundation (2025). Model Context Protocol Specification 2025-11-25. (Tool integration protocol)

Full research background, citations, and methodology: docs/research/

Why AIWG

For Individual Developers

Turn your AI coding assistant from a stateless autocomplete into a project-aware development partner. Without AIWG, every time your AI assistant restarts you lose all context. With AIWG, the .aiwg/ directory maintains requirements, architecture decisions, test strategies, and project history across sessions. The Ralph loop means you can hand off complex multi-step tasks ("migrate this module to TypeScript") and walk away — the agent iterates until completion or escalates when stuck.

For Engineering Teams

Standardize how your team works with AI across 8 platforms. Whether your team uses Claude Code, Copilot, Cursor, or Warp, everyone gets the same agents, rules, and workflows. The 35 enforcement rules prevent common AI mistakes: deleting tests to make them pass, fabricating citations, hard-coding tokens, or silently dropping features. Human-in-the-loop gates at phase transitions ensure no AI output reaches production without human review.

For Platform Engineers

Deploy consistent AI-augmented workflows across your organization. AIWG's extension system lets you create custom agents, commands, and skills specific to your domain, then deploy them to any supported platform. The scaffolding commands (aiwg add-agent, aiwg scaffold-addon, aiwg scaffold-framework) make it easy to build and distribute organizational capabilities.

For Researchers

Standards-aligned implementation of multi-agent systems with full provenance. AIWG operationalizes FAIR Principles (G20, EU, NIH endorsement), W3C PROV for provenance tracking, and GRADE methodology for evidence quality. The research framework manages paper discovery, citation integrity, and archival lifecycles. All 138+ research paper citations are grounded in verified sources — no hallucinated references.

For Security Teams

Forensics-grade investigation workflows and security review automation. The Forensics Complete framework provides 13 specialized agents for digital forensics and incident response, with NIST SP 800-86 evidence handling, MITRE ATT&CK mapping, Sigma rule hunting, and STIX 2.1 IOC formatting. Security review cycles integrate into the SDLC with automated threat modeling and vulnerability management.

AIWG vs Manual AI Workflows

Standards & Compliance

Documentation

Getting Started

• Quick Start Guide — Install and deploy in minutes
• Prerequisites — Node.js, AI platforms, OS support
• CLI Reference — All 47 commands with examples

By Audience

Practitioners:

• Quick Start Guide — Hands-on workflows
• Ralph Loop Guide — Iterative execution with crash recovery
• Platform Guides — 5-10 minute setup per platform

Technical Leaders:

• Extension System Overview — Architecture and capabilities
• Workspace Architecture — Multi-framework isolation
• Multi-Agent Orchestration — Ensemble patterns

Researchers & Evaluators:

• Research Background — Literature review and citations
• Glossary — Professional terminology mapping
• Production-Grade Guide — Failure mode mitigation

Platform Guides

• Claude Code — 5-10 min setup
• Warp Terminal — 3-5 min setup
• Factory AI — 5-10 min setup
• Cursor — 5-10 min setup
• All Integrations

Framework Documentation

• SDLC Framework — 90 agents, phase workflows, quality gates
• Forensics Complete — DFIR investigation workflows
• Marketing Kit — 37 agents, campaign lifecycle
• Media Curator — Media archive management
• Research Complete — 8-stage research pipeline

Extension System

AIWG's unified extension system enables dynamic discovery, semantic search, and cross-platform deployment:

• Extension System Overview — Architecture and capabilities
• Creating Extensions — Build custom agents, commands, skills
• Extension Types Reference — Complete type definitions

Advanced Topics

• Ralph Loop — Iterative task execution with crash recovery
• RLM Addon — Recursive context decomposition for 10M+ token processing
• Daemon Mode — Background file watching, cron scheduling, IPC
• Messaging Integration — Bidirectional Slack, Discord, and Telegram bots
• MCP Server — Model Context Protocol integration
• Agent Design Bible — 10 Golden Rules for agent creation
• YAML Metalanguage — Declarative workflow schemas

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Quick contributions:

• Found an AI pattern? Open an issue
• Have a better rewrite? Submit a PR to examples/
• Want to add an agent? Use aiwg add-agent or see docs/development/agent-template.md
• Want to add a skill? Use aiwg add-skill
• Want to create an addon? Use aiwg scaffold-addon

Community & Support

• Website: aiwg.io
• Discord: Join Server
• Telegram: Join Group
• Issues: GitHub Issues
• Discussions: GitHub Discussions

Usage Notes

AIWG is optimized for token efficiency. Rules deploy as a consolidated index (~200 lines) instead of 35 individual files (~9,321 lines). Most users on Claude Pro or similar plans will have no issues. See Usage Notes for rate limit guidance.

License

MIT License — Free to use, modify, and distribute. See LICENSE.

Important: This framework does not provide legal, security, or financial advice. All generated content should be reviewed before use. See Terms of Use for full disclaimers.

Sponsors

Interested in sponsoring? Contact us

Acknowledgments

Research foundations: Built on established principles from cognitive science (Miller 1956, Sweller 1988), multi-agent systems (Jacobs et al. 1991, MetaGPT, AutoGen), software engineering (Cooper 1990, RUP), and recent AI systems research (ReAct, Self-Refine, DSPy, SWE-Agent). Implements standards from FAIR Principles, OAIS (ISO 14721), W3C PROV, GRADE evidence assessment, and MCP protocol (Linux Foundation).

Platforms: Thanks to Anthropic (Claude Code), GitHub (Copilot), Warp, Factory AI, Cursor, and the OpenCode community for building the platforms that enable this work.