HarnessCommunity★ 1k

MetaSwarm

Updated Jul 7, 2026

A self-improving multi-agent orchestration framework by dsifry extracted from a production SaaS codebase. 18 specialized agent personas, a 9-phase workflow (Research → Plan → Design Review Gate → Work Unit Decomposition → Orchestrated Execution → Final Review → PR Creation → PR Shepherd → Closure & Learning), BEADS git-native task tracking, and a JSONL knowledge base that grows with every PR. Works with Claude Code, Gemini CLI, and Codex CLI.

View on GitHub

What it does

/start-task
Launches the 9-phase workflow from a plain-English description; routes through Research, Plan, Design Review Gate, and into Orchestrated Execution.
/setup
Interactive project setup that configures CLAUDE.md, AGENTS.md, knowledge base templates, and BEADS task tracking.
Design Review Gate
Five specialist agents (PM, Architect, Designer, Security, CTO) review the design in parallel; blocks decomposition until the gate passes or escalates to human after 3 iterations.
Orchestrated Execution Loop
Per-task 4-phase loop (IMPLEMENT → VALIDATE → ADVERSARIAL REVIEW → COMMIT) where the orchestrator validates independently and a different model reviews the writer's work.
/self-reflect
Post-PR reflection that extracts patterns, gotchas, and anti-patterns from reviewer comments and session history, writing them to the JSONL knowledge base.
npx metaswarm init
Cross-platform installer that detects installed AI CLIs and installs MetaSwarm for all of them in one pass.

Files (1)

OVERVIEW.mdprimary · markdown · 2.4 KB

## What it does

MetaSwarm is a production-tested multi-agent orchestration framework extracted from a SaaS codebase with 100% test coverage and hundreds of autonomous PRs. It provides:

18 agent personas: Researcher, Architect, Coder, Security Auditor, PR Shepherd, and others.

9-phase workflow: Research → Plan → Design Review Gate → Work Unit Decomposition → Orchestrated Execution (4-phase loop per task: IMPLEMENT → VALIDATE → ADVERSARIAL REVIEW → COMMIT) → Final Review → PR Creation → PR Shepherd → Closure & Learning.

Parallel Design Review Gate: 5 specialist agents (PM, Architect, Designer, Security, CTO) review in parallel with a 3-iteration cap before human escalation. The gate blocks decomposition until it passes.

Orchestrated Execution Loop: orchestrator validates independently (runs tests itself, never trusts subagent self-reports), adversarial reviewers check DoD compliance with file:line evidence, and optionally delegates IMPLEMENT or REVIEW to external models (Codex CLI, Gemini CLI) for cost savings and cross-model adversarial review.

BEADS task tracking: git-native issue/task management with dependency tracking and knowledge priming. Selective retrieval (bd prime --files ... --keywords ... --work-type ...) loads only the relevant knowledge subset before each task.

Self-learning: after every PR merge, /self-reflect analyzes reviewer comments, build failures, and architectural decisions, writing structured JSONL entries back to the knowledge base. The system also introspects the session for repeated user corrections and friction points, proposing new skills or commands.

Supports Claude Code (plugin marketplace), Gemini CLI (extension), and Codex CLI (plugin marketplace). Cross-platform installer: npx metaswarm init.

## Install

Claude Code:

claude plugin marketplace add dsifry/metaswarm-marketplace
claude plugin install metaswarm

Then:

/setup

Start a task:

/start-task Add a webhook system with retry logic, signature verification, and a delivery log UI.

## When to use

MetaSwarm is designed for team-scale AI development on production codebases where quality gates must be non-negotiable. The adversarial review pattern (writer always reviewed by a different model or agent) and mandatory TDD coverage thresholds mean it's slower than single-shot prompting but catches integration breakage before it reaches a human reviewer. Best suited for features that cross multiple modules or require security review.