Sub-Agents: Coordinated Multi-Agent Workflows for Claude Code

We built a Node.js CLI that extends Claude Code from a single-agent assistant into a coordinated development team. Sub-Agents Manager ships 60+ pre-built specialist agents — backend architects, security auditors, data scientists, infrastructure engineers, documentation writers, and more — organized across 14 functional domains, with workflow orchestration that runs groups of agents concurrently and coordinates their findings through shared persistent memory.

Installation is a single command:

npm install -g @webdevtoday/claude-agents
claude-agents init
claude-agents workflow --template fullstack-development

Agents placed in ~/.claude/agents/ or .claude/agents/ are automatically discovered by Claude Code as native sub-agents, available in chat without any additional configuration.

Why specialized agents over a single generalist

A single generalist agent making sequential decisions across a full-stack project loses context between tasks and cannot run multiple workstreams at the same time. Running five specialist agents sequentially produces the same second problem, plus another: manual context passing between stages.

The workflow model in Sub-Agents addresses both. Agents within a phase run concurrently via Promise.all(). Measured tests show a 79.7% performance improvement for four-agent parallel execution versus sequential. A seven-phase fullstack development workflow runs Architecture & Planning (backend-architect, frontend-developer, project-planner — parallel), then Backend Development (api-developer, tdd-specialist — parallel), through QA, DevOps, Performance, and Documentation.

Specialization also improves output quality. A backend architect agent carries a deep system prompt shaped by 15 years of architecture context. A TDD specialist reasons differently about the same codebase than a security auditor does.

How team coordination and memory work

The coordination backbone is SimpleMemoryStore — a namespaced, TTL-aware, disk-backed JSON store at .swarm/memory.json. Every agent result is automatically written under a structured key:

workflow:<id>:agent:<agent-name>:phase:<n>:result

Downstream agents retrieve prior findings without any manual context passing:

// A synthesis agent in phase 3 reads all prior findings:
const priorWork = memory.getByPattern(`workflow:${workflowId}:agent:*`);

The store auto-saves every 30 seconds and survives process restarts.

Quality gates are declared in YAML specification files, not embedded in agent prompts. This separates policy from mechanics: the criterion test coverage >= 80% lives in a versioned, auditable document. A phase does not advance if its gate fails.

Agent definitions are Markdown files with YAML frontmatter — version-controllable, human-readable, and portable. Moving an agent between user scope, project scope, or an npm package requires no modification. Later scopes override earlier ones.

Sub-Agents also automatically detects context-forge projects. If a project contains CLAUDE.md, a PRPs/ directory, or a .claude/commands/ folder, the system parses the project rules and injects that context into every agent execution — with no configuration required.

Where it applies

Sub-Agents is the agent library and orchestration layer sitting between Claude Code and the Worksona API. Where the Worksona Delegator handles delegation patterns, Sub-Agents provides the 60+ specialists that can be assembled into those patterns. The dependency count is intentionally minimal — 7 production packages, no external database, no external API calls in core — keeping the system portable and straightforward to extend.