Gem Team is a self-learning multi-agent orchestration harness designed for spec-driven development and automated verification.

Performance

• 4x Faster — Parallel execution with wave-based execution
• Pattern Reuse — Codebase pattern discovery prevents reinventing wheels

Quality & Security

• Higher Quality — Specialized harness agents + TDD + verification gates + contract-first
• Built-in Security — OWASP scanning, secrets/PII detection on critical tasks
• Resilient — Pre-mortem analysis, failure handling, auto-replanning
• Accessibility-First — WCAG compliance validated at spec and runtime layers
• Safe DevOps — Idempotent operations, health checks, mandatory approval gates
• Constructive Critique — gem-critic challenges assumptions, finds edge cases

Intelligence

• Established Patterns — Uses library/harness conventions over custom implementations
• Source Verified — Every factual claim cites its source; no guesswork
• Knowledge-Driven — Prioritized sources (PRD → codebase → AGENTS.md → Context7 → docs)
• Continuous Learning — Memory tool persists patterns, gotchas, user preferences across sessions
• Auto-Skills — Agents extract reusable SKILL.md files from successful tasks (high confidence: auto, medium: confirm)
• Skills & Guidelines — Built-in skill & guidelines (web-design-guidelines)
• Context7 Integration — Real-time library documentation via Context7

Process

• Spec-Driven — Multi-step refinement defines "what" before "how"
• Verified-Plan — Complex tasks: Plan → Verification → Critic
• Traceable — Self-documenting IDs link requirements → tasks → tests → evidence
• Intent vs. Compliance — Shifts the burden from writing "perfect prompts" to enforcing strict, YAML-based approval gates
• Diagnose-then-Fix — gem-debugger diagnoses → gem-implementer fixes → re-verifies
• Pre-Mortem — Failure modes identified BEFORE execution
• Contract-First — Contract tests written before implementation

Token Efficiency

Optimized for reduced LLM token consumption without quality loss:

• Concise Output — No preamble, no meta commentary, no verbose explanations
• Strict Formats — JSON/YAML exactly matching schemas — eliminates parse errors and retries
• Empty is OK — Skip empty arrays, nulls, verbose fields where not needed
• File-Based — Researcher/Planner save to YAML files (not all in JSON output)
• Learnings — Empty patterns/conventions unless critical

Result: ~40-60% reduction on output tokens while maintaining quality.

Design

• Design Agents — Dedicated agents for web and mobile UI/UX with anti-"AI slop" guidelines for distinctive aesthetics
• Mobile Agents — Native mobile implementation (React Native, Flutter) + iOS/Android testing

Triple Learning System

Harness Architecture

User Goal → Orchestrator → [Simple: Research/Plan] or [Complex: Discuss → PRD → Research → Plan → Approve] → Execute (waves) → Summary → Final Review
                ↓
            Diagnose → Fix → Re-verify

Next Steps

• Read about the Agent Team
• Learn about the Installation process
• Explore the Contributing guidelines