Agentic AI Atlasby a5c.ai
OverviewWikiGraphFor AgentsEdgesSearchWorkspace
/
GitHubDocsDiscord
iiRecord
Agentic AI Atlas · MetaSwarm
specialization:metaswarma5c.ai
Search record views/
Record · tabs

Available views

II.Record viewspp. 1 - 1
overviewarticlejsongraph
III.Related pagespp. 1 - 1
II.
Specialization reference

specialization:metaswarm

Reading · 4 min

MetaSwarm reference

Metaswarm is an autonomous multi-agent orchestration framework that manages the complete lifecycle from GitHub issue to merged pull request. It coordinates 12 specialized agents through a rigorous 7-phase workflow with quality gates as blocking state transitions, adversarial reviews with fresh reviewers, and knowledge persistence across sessions.

Specializationwiki/library/metaswarm.mdOutgoing · 1Incoming · 2

Metaswarm Methodology

**Source**: dsifry/metaswarm by David Sifry **Category**: Autonomous Multi-Agent Orchestration / Issue-to-PR Lifecycle **License**: See upstream repository

Overview

Metaswarm is an autonomous multi-agent orchestration framework that manages the complete lifecycle from GitHub issue to merged pull request. It coordinates 12 specialized agents through a rigorous 7-phase workflow with quality gates as blocking state transitions, adversarial reviews with fresh reviewers, and knowledge persistence across sessions.

Core Principles

  • **Trust Nothing, Verify Everything, Review Adversarially** - Quality gates are blocking, never advisory
  • **TDD Mandatory** - Write tests first, watch them fail, then implement (100% coverage targets)
  • **Fresh Reviewer Rule** - On re-review after FAIL, spawn new reviewer with no memory (prevents anchoring bias)
  • **Independent Validation** - Orchestrator runs quality gates directly, never trusts subagent self-reports
  • **Human Checkpoints** - Planned pauses at critical boundaries (schema changes, security code, new patterns)
  • **Knowledge Persistence** - Extract learnings while context is fresh, persist for cross-session continuity

Process Files

ProcessFileDescriptionTask Count
Issue Orchestratormetaswarm-orchestrator.jsFull 7-phase lifecycle: research to PR12
Design Review Gatemetaswarm-design-review.js6 parallel specialist reviews, unanimous approval7
Execution Loopmetaswarm-execution-loop.js4-phase cycle: Implement -> Validate -> Review -> Commit4
Swarm Coordinatormetaswarm-swarm-coordinator.jsMulti-issue parallel management across worktrees6
Knowledge Cyclemetaswarm-knowledge-cycle.jsContext priming and self-reflection5
PR Shepherdmetaswarm-pr-shepherd.jsPR lifecycle through merge3

Skills Catalog

SkillDirectoryDescription
orchestrated-executionskills/orchestrated-execution/4-phase execution loop with quality gates
design-review-gateskills/design-review-gate/6-agent parallel design review
plan-review-gateskills/plan-review-gate/3 adversarial plan reviewers
adversarial-reviewskills/adversarial-review/Fresh reviewer with binary PASS/FAIL
knowledge-curationskills/knowledge-curation/Context priming and self-reflection
work-unit-decompositionskills/work-unit-decomposition/DoD items, file scope, dependencies
pr-shepherdingskills/pr-shepherding/PR lifecycle management through merge
external-tool-coordinationskills/external-tool-coordination/Cross-model AI tool integration

Agents Catalog

AgentDirectoryRole
issue-orchestratoragents/issue-orchestrator/Master coordinator per issue
researcheragents/researcher/Codebase exploration and analysis
architectagents/architect/Implementation planning and decomposition
product-manageragents/product-manager/Use case and scope validation
designeragents/designer/UX/API design review
security-designagents/security-design/Threat modeling and OWASP analysis
ctoagents/cto/TDD readiness and codebase alignment
coderagents/coder/TDD implementation specialist
code-revieweragents/code-reviewer/Fresh adversarial reviewer
security-auditoragents/security-auditor/Implementation security review
pr-shepherdagents/pr-shepherd/PR lifecycle management
swarm-coordinatoragents/swarm-coordinator/Multi-issue parallel orchestration

Workflow Lifecycle

Code
Issue -> Research -> Plan -> Plan Review Gate (3 adversarial) -> Preflight -> Design Review Gate (6 parallel, unanimous) -> Work Unit Decomposition -> Execution Loop (Implement -> Validate -> Adversarial Review -> Commit) x N -> Final Review -> Self-Reflect -> PR -> Shepherd -> Merge

Quality Gates (Blocking)

1. **Plan Review Gate** - 3 adversarial reviewers (Feasibility, Completeness, Scope) 2. **Design Review Gate** - 6 parallel specialists (ALL must approve, max 3 iterations) 3. **Quality Gate Validation** - Independent tsc, eslint, vitest (blocking, never advisory) 4. **Adversarial Code Review** - Fresh reviewer, binary PASS/FAIL (max 3 attempts -> escalate) 5. **Coverage Gate** - 100% target across lines/branches/functions/statements

Anti-Patterns (Enforced)

  • Self-certifying (trusting subagent claims)
  • Combining execution phases into single steps
  • Reusing reviewers after FAIL verdict
  • Passing previous review findings to new reviewers
  • Treating quality gate failures as advisory
  • Proceeding past human checkpoints without explicit approval
  • Using --no-verify on commits or force-pushing

Philosophy

  • **Adversarial over collaborative** review for verified compliance
  • **Independent validation** over trusted reporting
  • **Blocking gates** over advisory recommendations
  • **Fresh reviewers** over experienced reviewers (prevent bias)
  • **Knowledge extraction** before context is lost
  • **Human escalation** after bounded retry (3 attempts)

Article source

Metaswarm Methodology (Library)

This record inherits its article from a related Page node.

Related pages

Metaswarm Methodology (Library)

Shortcuts

Open overview
Open JSON
Open graph