Help us improve
Share bugs, ideas, or general feedback.
From devflow
This directory contains custom subagents for the DevFlow autonomous development workflow.
npx claudepluginhub joshuarweaver/cascade-code-devops-misc-1 --plugin ao-cyber-systems-devflow-claudeHow this agent operates — its isolation, permissions, and tool access model
Agent reference
devflow:agents/readmeThe summary Claude sees when deciding whether to delegate to this agent
This directory contains custom subagents for the DevFlow autonomous development workflow. | Agent | Purpose | Memory | Mode | |-------|---------|--------|------| | `trd-implementer` | Implements a single TRD | project | acceptEdits | | `trd-designer` | Designs TRDs by exploring codebase | project | plan (read-only) | | `code-reviewer` | Reviews code quality and security | project | plan (read-o...
Guide for Claude Code subagents: explains architecture, delegation model, creation via Markdown YAML frontmatter in .claude/agents/, invocation (@guide or auto), and configuration for tasks like code review, debugging, data analysis.
Implements subagent architecture for Claude Agent SDK applications. Handles orchestrator-specialist patterns, pipeline workflows, hierarchical task decomposition, and debate/consensus systems for multi-agent coordination.
Creates, improves, and analyzes Claude Code agents following Anthropic best practices. Delegate for explicit requests like 'create agent', 'new agent', 'improve agent', 'fix agent trigger', or 'agent doesn't trigger'.
Share bugs, ideas, or general feedback.
This directory contains custom subagents for the DevFlow autonomous development workflow.
| Agent | Purpose | Memory | Mode |
|---|---|---|---|
trd-implementer | Implements a single TRD | project | acceptEdits |
trd-designer | Designs TRDs by exploring codebase | project | plan (read-only) |
code-reviewer | Reviews code quality and security | project | plan (read-only) |
debugger | Investigates and fixes failures | project | acceptEdits |
Subagents run in isolated context windows, separate from the main conversation:
Main Conversation (Orchestrator)
├── TaskList, TaskCreate, TaskUpdate
├── High-level coordination
└── Spawns subagents for actual work
│
├── trd-implementer (isolated context)
│ ├── Reads TRD
│ ├── Implements feature
│ ├── Runs verification
│ └── Returns summary only
│
└── code-reviewer (isolated context)
├── Reviews changes
└── Returns findings only
Benefits:
Each subagent uses native Claude Code memory (not Memory MCP):
# In subagent frontmatter
memory: project # Persists in .claude/agent-memory/<agent-name>/
Memory locations:
project: .claude/agent-memory/<agent-name>/ (git-trackable)user: ~/.claude/agent-memory/<agent-name>/ (personal, all projects)local: .claude/agent-memory-local/<agent-name>/ (gitignored)MEMORY.md: First 200 lines auto-loaded into subagent's system promptFrom the orchestrator:
Task(subagent_type: "trd-implementer"):
Implement the TRD at .devflow/trds/TRD-001-feature.md
Follow all acceptance criteria.
The subagent:
agents/your-agent.md---
name: your-agent # Required: lowercase, hyphens
description: What this agent does # Required: used for auto-delegation
tools: Read, Edit, Bash # Optional: allowed tools
model: inherit # Optional: sonnet, opus, haiku, inherit
permissionMode: default # Optional: default, acceptEdits, plan, dontAsk
memory: project # Optional: project, user, local
skills: # Optional: preload skills
- verify
- regression
hooks: # Optional: agent-specific hooks
PreToolUse:
- matcher: "Bash"
hooks:
- type: command
command: "./validate.sh"
---
If you were using Memory MCP (mcp__memory__* tools):
Before (Memory MCP):
mcp__memory__create_entities: [...]
mcp__memory__search_nodes: "keyword"
After (Native Memory):
# In subagent frontmatter
memory: project
Subagent can now:
.claude/agent-memory/<name>/MEMORY.md (auto-loaded)The Memory MCP is still available but deprecated. Native subagent memory provides better integration with Claude Code's architecture.