A LangGraph-orchestrated AI development team with a SvelteKit dashboard for real-time monitoring. Three specialized agents (Architect, Lead Dev, QA) collaborate through structured JSON blueprints, execute code in E2B sandboxes, and persist knowledge via Chroma tiered memory.
agent-dev/
├── dev-suite/ # Python orchestrator (LangGraph)
│ ├── src/
│ │ ├── agents/ # Architect, Lead Dev, QA agent definitions
│ │ ├── api/ # FastAPI backend (REST + SSE)
│ │ ├── memory/ # Chroma vector store with tiered metadata
│ │ ├── sandbox/ # E2B sandbox runner
│ │ ├── tools/ # MCP bridge and tool providers
│ │ ├── orchestrator.py # LangGraph state machine
│ │ ├── cli.py # CLI interface
│ │ └── tracing.py # Langfuse observability
│ ├── tests/ # Comprehensive test suite
│ ├── pyproject.toml # uv/PEP 735 dependencies
│ └── mcp-config.json # MCP server version pins
├── dashboard/ # SvelteKit frontend
│ ├── src/
│ │ ├── lib/ # Stores, SSE client, context
│ │ └── routes/ # Layout + page components
│ ├── package.json # pnpm dependencies
│ └── svelte.config.js
├── .github/ # Issue templates, labels, CI workflows
├── CLAUDE.md # Claude Code context
└── CONTRIBUTING.md # Contribution guidelines
Orchestrator: LangGraph state machine with explicit transitions. Three agents collaborate in a plan → build → test loop with structured JSON blueprints, max 3 retries per task, and human escalation on budget exhaustion.
Agent Team:
| Role | Model | Responsibility |
|---|---|---|
| Architect | Gemini 2.5 Flash | Creates structured blueprints. Never writes code. |
| Lead Dev | Claude Sonnet 4 | Executes blueprints. Writes and refactors code. |
| QA Agent | Claude Sonnet 4 | Runs tests, audits security, writes failure reports. |
Memory: Chroma with tiered metadata (L0-Core human-only, L0-Discovered agent-writable with 48h expiry, L1 module context, L2 ephemeral).
Execution: E2B sandboxed micro-VMs with structured JSON output wrappers. Role-specific sandbox profiles (locked-down for Dev/QA, permissive for research).
Dashboard: SvelteKit (Svelte 5 + TailwindCSS v4) with SSE real-time streaming. VS Code-inspired layout with activity bar, sidebar panels, and bottom terminal. Degrades gracefully when backends are unavailable.
cd dev-suite
uv sync # Install dependencies
cp .env.example .env # Configure API keys
uv run python -m src # Run orchestrator
uv run pytest tests/ -v # Run testscd dev-suite
uv sync --group api
uv run --group api uvicorn src.api.main:app --reload --port 8000API docs at http://localhost:8000/docs.
cd dashboard
pnpm install
cp .env.example .env # Set BACKEND_URL
pnpm dev # http://localhost:5173| Method | Path | Description |
|---|---|---|
GET |
/health |
Health check (no auth) |
GET |
/agents |
Agent status list |
GET |
/tasks |
Task list with timelines |
GET |
/tasks/{id} |
Task detail with blueprint |
POST |
/tasks |
Create new task |
POST |
/tasks/{id}/cancel |
Cancel running task |
POST |
/tasks/{id}/retry |
Retry failed task |
GET |
/memory |
Memory entries (filterable) |
PATCH |
/memory/{id} |
Approve/reject memory |
GET |
/prs |
Pull request list |
GET |
/events |
SSE stream |
Set in dev-suite/.env:
ANTHROPIC_API_KEY— Claude API accessGOOGLE_API_KEY— Gemini API accessE2B_API_KEY— Sandbox executionLANGFUSE_PUBLIC_KEY/LANGFUSE_SECRET_KEY— Observability (optional)API_SECRET— Dashboard API auth token
Set in dashboard/.env:
BACKEND_URL— API base URL (e.g.http://localhost:8000)API_SECRET— Must match the API's secret
- ✅ LangGraph orchestrator with 3-agent team
- ✅ E2B sandbox execution
- ✅ Chroma tiered memory
- ✅ MCP tool bridge (Filesystem + GitHub)
- ✅ FastAPI dashboard backend (REST + SSE)
- ✅ SvelteKit dashboard scaffold (VS Code layout)
- ✅ GitHub Actions CI
- 🚧 Dashboard data wiring (stores ↔ API)
- 🚧 Memory approval UI (L0-Discovered review)
- 🚧 OTEL trace propagation
- 📋 CI/CD Pipeline MCP
- 📋 Secrets provider integration
- 📋 Cost alerting thresholds