"Explore once. Exploit next."

An agentic harness where agents learn from trial-and-error execution and codify learnings into reusable skills.

Why Tsugi?

Most agent memory stores facts. Most agent skills are static. Tsugi brings the best from the two worlds.

What It Does

• Run 1: Research + Task Execution + Skill Creation
• Run 2: Skill Lookup + Skip Research (faster and token saving)

Tsugi captures procedural knowledge from successful task executions and saves them as reusable skills. When you ask an agent to integrate with a new API, it researches, experiments, handles errors, and eventually succeeds. That hard-won knowledge gets codified into a skill that makes the next execution instant.

Skills encode two types of knowledge:

• Procedural: Integration gotchas, validation rules, error patterns, multi-step workflows
• Preferences: Your taxonomies, classification rules, domain constraints

Features

• Dual-Agent System - Task Agent executes work, Skill Agent codifies learnings
• Persistent Skill Library - Skills accumulate and compound over time
• Sandbox Execution - Isolated environments with environment variable injection
• Real-time Streaming - Watch agent reasoning and tool calls as they happen
• Extended Thinking - Full visibility into agent reasoning traces
• Native Grounding - Google Search and URL analysis built-in
• Conversation History - Persistent chat history with pinned comparisons

How It Works

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│   User Task     │ ──▶ │   Task Agent    │ ──▶ │  Skill Agent    │
│                 │     │                 │     │                 │
│ "Charge $50     │     │ 1. Search skills│     │ 1. Analyze      │
│  on Stripe"     │     │ 2. Research API │     │    transcript   │
│                 │     │ 3. Execute      │     │ 2. Extract      │
│                 │     │ 4. Verify       │     │    procedure    │
│                 │     │ 5. Suggest      │     │ 3. Save skill   │
│                 │     │    codification │     │                 │
└─────────────────┘     └─────────────────┘     └─────────────────┘

First execution - Agent explores, makes mistakes, self-corrects, and succeeds.

"Codify Skill" - Skill Agent extracts the working procedure with parameters.

Next execution - Agent finds the skill, skips research, executes directly.

Tech Stack

• Framework: Next.js 16 with React 19 (App Router)
• LLM: Gemini 3 via Vercel AI SDK
• Database: SQLite/Turso for conversations, Vercel Blob for skills (prod)
• Sandbox: Local child_process (dev) or Vercel Sandbox microVM (prod)
• Observability: Braintrust for traces and token counting
• Styling: Tailwind CSS v4, Framer Motion, Lucide icons

Getting Started

Prerequisites

• Node.js 18+
• pnpm

Setup

# Install dependencies
pnpm install

# Configure environment
cp .env.example .env

Add your API key to .env:

GOOGLE_GENERATIVE_AI_API_KEY=your_key_here

Get a key from Google AI Studio.

Run

pnpm dev

Open http://localhost:3000.

Project Structure

src/
├── app/                    # Next.js App Router
│   ├── api/agent/          # SSE streaming endpoint
│   ├── api/conversations/  # Chat history CRUD
│   ├── api/skills/         # Skills API
│   └── task/               # Task execution page
├── components/             # React components
│   └── landing/            # Landing page components
├── hooks/                  # useForgeChat, useConversations, useSkills
└── lib/
    ├── agent/              # Task Agent, Skill Agent, tools
    ├── db/                 # SQLite/Turso database
    ├── sandbox/            # Sandbox executors
    ├── skills/             # Skill storage (local/cloud)
    └── tools/              # Command execution
playground/                 # Demo tasks
MEMORY/                     # Plans and changelogs

Development

pnpm dev        # Start dev server
pnpm build      # Production build
pnpm test       # Run tests (watch mode)
pnpm test:run   # Run tests once
pnpm lint       # Lint code

Example Tasks

Try these to see the system in action:

• "Send a hello world message to Discord" (requires webhook URL)
• "Charge $50 via Stripe API and then refund half" (requires Stripe key)
• "Summarize this YouTube video and save to Notion"

After successful execution, click "Codify Skill" to save the procedure. Run the same task again to see the speedup.

License

MIT