Gemini 3 Hackathon Entry

Explore once. Exploit next.

An agent harness that turns trial-and-error into reusable skills.

Don't have a key? Get one free from Google AI Studio

tsugi.exe

Run 1: Exploration

Trial-and-error, research, learning.

0.0s

Ready to start...

Run 2: Exploitation

Direct execution, no research.

0.0s

Waiting for optimized skill...

Never waste an experience.

tsugi enables an agent to memorize procedures, not facts.

STEP 01

Execute

The agent attempts a task out of the box. It might detour and self-correct.

STEP 02

Learn

Upon success, tsugi analyzes the trace. It strips away the dead ends and identifies the critical path.

STEP 03

Reuse

The successful procedure is codified into a skill. Next time, the agent retrieves this skill.

STEP 04

Compound

Agents build a library of executable capabilities that grows over time.

BUILT WITH

Gemini 3
Core LLM
Google AI Studio
Design
Vercel
AI SDK & Hosting
Claude Code
Development
Braintrust
Observability