# The AI4SD Open Source Landscape: From SDD to the Outer Loop
The AI software delivery open-source ecosystem is organized into six layers: SDD specs, Harness/Skills, context orchestration, task decomposition, agent platforms, and Outer Loop CI/CD — pick a starting point based on what you mainly want to control.
The AI4SD (AI Software Delivery) open-source ecosystem is large and evolving fast. This list is organized by research purpose, to help clarify how Spec-Driven Development, Harness, Agent Runtimes, and the Outer Loop relate to one another — it is not an exhaustive catalog, and star counts are only a rough popularity signal.
Positioning within the Three Loops
Inner Loop (AI pair-programming) → Middle Loop (Spec/Harness/Review) → Outer Loop (PR·CI·CD·Metrics)See AI4SE End-to-End Best Practices for the full E2E methodology, and AI Development Workflows Landscape for a comparison of workflow projects.
1. Spec-Driven Development (SDD)
How do you structure “what to build” into a spec an AI can execute?
| Project | Positioning | Features |
|---|---|---|
| GitHub Spec Kit | Official SDD standard flow | constitution → specify → plan → tasks → implement; 30+ agents |
| OpenSpec | Brownfield-first | Delta specs, change isolation; fluid not rigid |
| BMAD Method | Multi-role agile | 12+ agent roles (PM/Architect/Dev/UX/QA) |
| GSD Core | Execution orchestration + SDD | Discuss→Plan→Execute→Verify→Ship; resists context rot |
See A Side-by-Side Comparison of SDD Tools for a deeper comparison.
2. Harness, Skills, and Engineering Methodology
How do you get an agent to follow an engineering process?
| Project | Positioning |
|---|---|
| Superpowers | Enforces TDD, planning, verification, code-review skills |
| Everything Claude Code (ECC) | Cross-platform harness: Agents + Skills + Hooks + Rules + MCP |
| Cursor Rules / Plan Mode | IDE-native harness (Plan, five-step TDD, worktrees) |
See The Nine Types of Claude Code Skills for skills in practice, and Harness Engineering for the theory.
3. Context Engineering and Long-Horizon Orchestration
| Resource/project | Key points |
|---|---|
| Anthropic Context Engineering | JIT retrieval, compaction, sub-agents |
| Anthropic P/G/E Harness | Sprint contracts, Playwright verification |
| GSD Core | File-based artifacts, fresh context per task |
4. Task Decomposition and PRD Orchestration
| Project | Positioning |
|---|---|
| Taskmaster AI | PRD → task DAG; MCP integration into IDEs |
| Spec Kit / OpenSpec tasks | Lightweight task decomposition integrated with spec docs |
5. Agent Platforms and Pair-Programming Tools
IDE / CLI Agents (selected)
| Type | Representatives |
|---|---|
| AI-native IDE | Cursor, Windsurf, Kiro |
| IDE extension | Copilot, Cline, Roo Code, Continue |
| CLI | Claude Code, Codex CLI, Gemini CLI, OpenCode, Aider |
| Autonomous platform | OpenHands, SWE-agent |
Boundary note: OpenHands is an execution platform, while OpenSpec is a workflow framework — they shouldn’t be compared directly.
Interoperability Protocols
| Protocol | Role |
|---|---|
| MCP | Agent ↔ tools/data sources |
| ACP | IDE ↔ Agent standard interface (JetBrains × Zed) |
See The MCP Protocol.
6. Outer Loop: CI/CD and Quality Gates
How does AI-written code close an org-level loop through PR → CI → deployment?
| Category | Representatives | Role |
|---|---|---|
| CI engine | GitHub Actions, GitLab CI, Jenkins, Tekton | Pipeline as Code |
| GitOps CD | Argo CD, Flux CD | Auditable releases |
| Progressive delivery | Argo Rollouts, Flagger | Canary/blue-green |
| SAST | Semgrep (with MCP) | Security scanning of AI-written code |
| E2E | Playwright | Browser-based verification evidence |
| MCP integration | GitHub MCP Server | Issues/PRs/Actions |
Consultant’s principle: pilots usually don’t replace the CI platform — they layer AI code gates and metrics on top of the existing Outer Loop.
Suggested Reading Order
- Build a mental picture of the three loops — E2E best practices + Inner/Middle/Outer Loop
- The SDD spectrum — Spec Kit → OpenSpec → SDD comparison
- Harness — Superpowers → Harness Engineering
- Context orchestration — GSD Core → Loop Engineering
- Outer Loop — GitHub Actions + Semgrep + Playwright
Quick Reference by Goal
| Research goal | Prioritize |
|---|---|
| SDD + three-loop pilot | §1 SDD, §2 Harness |
| Middle Loop design | §2, §3 Context |
| Outer Loop / release | §6 CI/CD, SAST, E2E |
| End-to-end autonomous agents | §5 OpenHands, SWE-agent |
| Tool-chain combinations | OpenSpec+Superpowers+gstack |
Disclaimer
This list includes open-source projects alongside some commercial/platform products (Cursor, Copilot, Claude Code, etc.) for comparison. Star counts, features, and licenses should be checked against each repo’s live data; enterprise selection should factor in team constraints, compliance, and task risk — not popularity alone.