Agent-Native-SDLC

Agent-Native SDLC (Software Factory Pattern)

Agent-Native SDLC is an architectural shift where the entire software development lifecycle—from requirements to deployment—is designed for collaboration between human stakeholders and autonomous AI agents.

🏭 The “Software Factory” Model

Unlike traditional “Single-Player” AI coding tools (which focus on the IDE), the Software Factory model focuses on the Orchestration Layer.

1. Upstream Decision Capture

The system captures the “Why” before the “How.”

• Requirements: Structured, agent-readable business intent.
• Architecture: Explicit trade-offs and component maps.
• Context Engineering: Bridging the gap between a PM’s document and an Engineer’s implementation.

2. The Knowledge Graph

At the core of an Agent-Native system is a Knowledge Graph that links every artifact:

• Requirement A → Architectural Decision B → Code Implementation C.
• Propagation: If Requirement A changes, the system identifies all impacted decisions and code, notifying agents to refactor or update context.

3. Specialized Agent Roles

Instead of one generalist agent, the system uses a “Symphony” of specialized agents:

• Architect Agent: Validates design constraints.
• Implementation Agent: Writes code based on structured plans.
• QA Agent: Generates tests against the original requirements.

At the individual session level, this pattern is implemented via Claude SubAgents: the orchestrator dispatches specialist subagents (Security Reviewer, Documentation Writer, Test Generator), each loaded with only the context relevant to their role.

🚀 Strategic Benefits

• Zero Tribal Knowledge: Decisions are documented in the graph, not just in Slack or heads.
• Coherence at Scale: Agents maintain a global view of the project that humans often lose as complexity grows.
• Synchronized Updates: Reduces “context drift” between documentation and reality.

MCP-Automated SDLC Pipeline

The GitHub MCP Server enables a fully automated SDLC pipeline within Claude Code:

Spec → Plan → Implement → Test → Review → [Commit → Push → Create PR → Merge → Delete Branch]
                                                              ↑
                                                    Automated via GitHub MCP

The final bracketed steps — traditionally manual git operations — become a single natural language prompt: “Commit all changes, push to feature branch, create PR into main, squash merge, switch to main, pull latest, delete feature branch.”

Prerequisite: GitHub Personal Access Token (PAT) with fine-grained permissions including PR write access. Without explicit PR permissions, the flow breaks at PR creation. See Claude + MCP Explained.

---

Source: Ingested from Introducing 8090
SubAgent pattern: CampusX: Claude SubAgents
Custom subagent configuration: CampusX: Claude Custom Subagents