Transform complex projects into coordinated, verified, production-ready deliverables — with intelligent agents that research, plan, implement, test, and document autonomously.
A modular, high-performance multi-agent team designed for complex project execution, feature implementation, and automated verification.
Traditional AI coding assistants hit walls when projects get complex:
- Context overload — One agent trying to hold everything leads to mistakes
- No specialization — Jack of all trades, master of none
- Sequential bottlenecks — Tasks execute one-by-one, wasting time
- Missing verification — Changes ship without proper testing
- No audit trail — What changed? Why? Who knows...
| Challenge | Gem Team Approach |
|---|---|
| 🧠 Context Overload | Specialized agents with focused expertise — each holds only what it needs |
| 🎯 Lack of Specialization | 8 expert agents: researcher, planner, implementer, tester, reviewer, devops, and documentation specialist |
| 🐢 Sequential Bottlenecks | DAG-based parallel execution — up to 4 agents work simultaneously |
| ❌ Missing Verification | Verification-first: no task completes without passing its verification command |
| 📜 No Audit Trail | Persistent plan.md state file tracks every decision, status, and outcome |
- 🚀 10x Faster Execution — Parallel agent execution eliminates bottlenecks
- 🎯 Higher Quality Output — Specialized agents + mandatory verification = fewer bugs
- 🔒 Built-in Security — Dedicated reviewer agent applies OWASP scanning on critical tasks
- 📊 Full Visibility — Real-time plan status, clear approval gates, comprehensive summaries
- 🔄 Resilient Workflows — Pre-mortem analysis, failure handling, and automatic replanning
Gem Team follows a Strategic Planner/Dynamic Orchestrator pattern. It decomposes high-level user goals into a Directed Acyclic Graph (DAG) of tasks, executes them in parallel across specialized agents, and maintains a rigorous state-controlled workflow.
┌─────────────────────────────────────────────────────────────────┐
│ USER GOAL │
└──────────────────────────────┬──────────────────────────────────┘
▼
┌──────────────────────────────────────────────────────────────────┐
│ 🎭 ORCHESTRATOR │
│ • Coordinates workflow • Delegates via runSubagent │
│ • Maintains plan.md • Synthesizes results │
└──────────────────────────────┬───────────────────────────────────┘
▼
┌──────────────────────┴──────────────────────┐
▼ ▼
┌───────────────────┐ ┌────────────────────────┐
│ 🔍 RESEARCHER │ ──────────────────▶ │ 📋 PLANNER │
│ Context Mapping │ findings │ DAG Task Decomposition│
└───────────────────┘ └────────────┬───────────┘
▼
┌────────────────────────┐
│ 📄 plan.md │
│ (Task DAG + State) │
└────────────┬───────────┘
▼
┌────────────────────┬─────────────────────────┼─────────────────────────┬────────────────────┐
▼ ▼ ▼ ▼ ▼
┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ 💻 IMPLEMENTER│ │ 🌐 CHROME │ │ ⚙️ DEVOPS │ │ 🛡️ REVIEWER │ │ 📝 DOC WRITER │
│ TDD Execution │ │ TESTER │ │ CI/CD + Infra│ │ Security Gate│ │ Technical │
└──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘
| Agent | Specialty | Primary Responsibility |
|---|---|---|
gem-orchestrator |
Coordination | Coordinates multi-agent workflows, delegates tasks, synthesizes results via runSubagent |
gem-researcher |
Research | Gathers codebase context, identifies relevant files/patterns, returns structured findings |
gem-planner |
Strategy | Creates DAG-based plans with pre-mortem analysis, presents for approval, iterates on feedback |
gem-implementer |
Execution | Executes TDD code changes, ensures verification, maintains quality |
gem-browser-tester |
Testing | Automates browser testing, UI/UX validation via Chrome DevTools |
gem-devops |
Infrastructure | Manages containers, CI/CD pipelines, and infrastructure deployment |
gem-reviewer |
Quality | Security gatekeeper — OWASP scanning, secrets detection, compliance |
gem-documentation-writer |
Knowledge | Generates technical docs, diagrams, maintains code-documentation parity |
flowchart TD
A[🎯 User Goal] --> B[🎭 Orchestrator]
B --> C[🔍 Researcher]
C --> D{⏸️ Findings Review}
D -->|Feedback| C
D -->|Approved| E[📋 Planner]
E --> F[📄 plan.md]
F --> G{⏸️ Plan Approval}
G -->|Feedback| F
G -->|Approved| H[🚀 Parallel Execution]
H --> I[💻 Implementer]
H --> J[🌐 Chrome Tester]
H --> K[⚙️ DevOps]
H --> L[🛡️ Reviewer]
H --> M[📝 Doc Writer]
I & J & K & L & M --> N[🔄 Synthesize Results]
N --> O{All Tasks Done?}
O -->|No| H
O -->|Yes| P[📊 Walkthrough Summary]
- Inception — Orchestrator receives goal → delegates to multiple Researchers (by focus area)
- Findings Review — Orchestrator presents consolidated research findings via
plan_review→ MANDATORY PAUSE → iterates on feedback until findings are approved - Planning — Planner synthesizes findings, creates 3-7 atomic tasks with dependencies, runs pre-mortem analysis, saves
plan.md - Plan Approval — Planner presents plan via
plan_review→ MANDATORY PAUSE → iterates on feedback until approved - Delegation — Orchestrator identifies "ready" tasks (dependencies met) → launches up to 4 agents in parallel via
runSubagent - Execution — Workers execute changes and run verification commands before returning results
- Synthesis — Orchestrator processes handoffs, updates
plan.md, routes failures for revision/retry - Batch Confirmation — Orchestrator presents summary via
walkthrough_review→ MANDATORY PAUSE - Loop — Repeat delegation/synthesis until all tasks complete. Triggers replanning if stuck.
- Delivery — Final comprehensive
walkthrough_reviewsummary presented
The Orchestrator identifies key domains or features and launches multiple Researcher agents in parallel, each targeting a specific focus_area. This ensures deep, specific context is gathered for every part of the system before the Planner synthesizes it all into a unified plan.md.
Up to 4 concurrent agents execute independent tasks simultaneously, dramatically reducing total execution time while maintaining resource stability.
No task completes without passing its defined verification command. Implementers follow strict TDD discipline:
- Write tests FIRST
- Confirm tests FAIL
- Write MINIMAL code to pass
- Check
get_errorsafter every edit
The Reviewer agent acts as a security gatekeeper for critical tasks:
- OWASP Top 10 scanning
- Secrets/PII detection
- Compliance verification
- Tiered review depth (Full → Standard → Lightweight)
For complex plans, the Planner runs pre-mortem analysis — identifying potential failure modes, their likelihood, impact, and mitigation strategies BEFORE execution begins.
State persists in docs/plan/{plan_id}/plan.md, providing:
- Recovery from interruptions
- Complex retry handling
- Clear audit trail of project evolution
- Full traceability of decisions
Agents share knowledge via a memory system with citations:
- Researcher reads memories to understand project context before exploration
- Planner stores architectural decisions, design patterns, tech stack choices
- Orchestrator stores project-level decisions, product vision, norms
- Just-in-time verification: Citations (file:line) are checked before using memories
User → Orchestrator → Subagents (via runSubagent)
- Orchestrator:
disable-model-invocation: true— delegates only, never executes - Subagents:
disable-model-invocation: false— execute tasks via tools - Isolation: Subagents cannot call other subagents — all collaboration mediated by Orchestrator
gem-team/
├── gem-orchestrator.agent.md # Coordination hub
├── gem-researcher.agent.md # Context gathering
├── gem-planner.agent.md # DAG-based planning
├── gem-implementer.agent.md # TDD code execution
├── gem-browser-tester.agent.md # Browser automation
├── gem-devops.agent.md # Infrastructure & CI/CD
├── gem-reviewer.agent.md # Security gatekeeper
├── gem-documentation-writer.agent.md # Technical docs
├── docs/
│ └── plan/{plan_id}/
│ ├── research_findings.yaml # Research output
│ └── plan.md # Task DAG state (markdown format)
└── README.md
| Scenario | How Gem Team Helps |
|---|---|
| Large Feature Implementation | Decomposes into parallel subtasks, implements with TDD, verifies each component |
| Codebase Refactoring | Researches patterns, plans migration, executes incrementally with tests |
| Security Audit | Reviewer scans for OWASP issues, secrets, compliance gaps |
| Documentation Overhaul | Doc Writer generates accurate docs maintaining code-documentation parity |
| CI/CD Pipeline Setup | DevOps agent creates containers, pipelines, deploys with health checks |
| UI/UX Testing | Chrome Tester automates validation matrix, captures visual evidence |
This project is licensed under the Apache License 2.0 — see the LICENSE file for details.
Built for Gem Team — Precision. Parallelism. Progress.
Transform complexity into coordinated execution.