Zing Forum

Reading

devt: Claude Code's Multi-Agent Development Workflow Plugin

devt is a lightweight multi-agent development workflow plugin designed for Claude Code. It achieves automatic complexity detection, closed-loop learning, memory layer management, and architecture health scanning through the collaborative work of 10 professional agents (programmers, testers, code reviewers, etc.).

Claude Code多智能体开发工作流AI辅助编程记忆层代码审查架构决策GraphifyRAG闭环学习
Published 2026-05-13 16:14Recent activity 2026-05-13 16:24Estimated read 5 min
devt: Claude Code's Multi-Agent Development Workflow Plugin
1

Section 01

devt: Claude Code's Multi-Agent Development Workflow Plugin (Core Overview)

devt (short for development team) is a lightweight multi-agent development workflow plugin designed for Claude Code. It decomposes traditional single AI programming sessions into a coordinated team of 10 professional agents (e.g., programmer, tester, code reviewer, architect). Key features include automatic complexity detection, closed-loop learning, memory layer management, and architecture health scanning. Unlike single large prompts, devt uses a "command → workflow → agent" architecture, with project rules stored in .devt/rules/ for language-agnostic generality.

2

Section 02

The Three Core Pain Points devt Solves

devt targets three main issues in AI-assisted programming:

  1. Session Amnesia: Single sessions forget architectural decisions after context window rollover—devt uses a permanent memory layer (/.devt/memory/) storing ADRs, concepts, flows, rejected proposals, and lessons.
  2. Surface Understanding: Single prompts often over-engineer simple fixes or ignore refactoring—devt auto-detects complexity levels (TRIVIAL/SIMPLE/STANDARD/COMPLEX) and applies corresponding pipelines.
  3. No Experience Accumulation: Session-end experience loss—devt implements a closed-loop learning cycle (extract lessons → curate → index → query → inject into context).
3

Section 03

devt's Core Components & Workflow Mechanisms

10 Professional Agents: Each agent (e.g., programmer, architect, retro) has dedicated roles and default skills (e.g., architect uses architecture-health-scanner). Memory Layer: Two tiers—temporary (/.devt/state/ for workflow-specific data) and permanent (/.devt/memory/ for规范 knowledge). Preflight Protocol: Two layers of protection:

  • Theme Brief: Auto-triggered to gather relevant project info (paths, symbols, rejected proposals, lessons).
  • File Guard: Checks edits against preflight rules before writing. Key Commands: /devt:do (auto-route tasks), /devt:workflow (build/improve features), /devt:debug (system debugging), /devt:council (decision pressure testing).
4

Section 04

Graphify Integration & Practical Setup

Graphify Benefits: For users with Graphify installed, devt gains:

  • AST parsing for accurate symbol matching (vs text grep).
  • Real impact analysis via import graph traversal.
  • Detection of stale symbols post-refactoring.
  • Token-efficient code search (10x savings). Installation:
  1. Add via Claude Code plugin marketplace: /plugin marketplace add emrecdr/devt/plugin install devt.
  2. Initialize: /devt:init (generates .devt configs).
  3. First workflow: /devt:workflow "add GET /health endpoint returning 200".
5

Section 05

devt's Significance for AI-Assisted Development

devt represents an evolution in AI-assisted development:

  1. From single sessions to continuous learning (cross-session experience accumulation).
  2. From general prompts to professional分工 (task-specific agents).
  3. From passive response to active protection (preflight protocol prevents errors).
  4. From personal tools to team collaboration (shared memory layer for project规范). It provides a落地, scalable framework for teams to adopt AI-assisted development while maintaining code quality and architectural consistency.