Sui Basecamp · Oct 7–8, 2026 · 30-min session

From binary to agents

A tour of how software gets built — and where AI takes it next.

Four layers of abstraction, each built on the one below it, ending somewhere new: a program you talk to that reasons, decides, and signs its own transactions. This page walks the whole path.

Start the tour Download the slides Sources

Before we start

A quick show of hands.

No wrong answers — this is just where the tour starts from for you. Tap the ones that apply.

Part 01

The evolution of software.

Every layer of software is built on the one beneath it. Here are four of them, stacked — from the machine at the bottom to plain language at the top. Open each one.

Each layer hides the one below it. Vibe-coding doesn't replace the others — it sits on top of them, the way Python sits on top of C, and C sits on top of binary.

“Maybe in ten years, talking to an AI will feel just as outdated. Maybe we'll just think, and a product appears.” — on where this evolution might go

Part 02

Building an agent.

A program that can reason, decide, and act — not just answer. It can call tools: search the web, run code, send a transaction.

Choosing your stack

Open-source · local

Run it on your machine

Qwen 2.5 (7B) or Hermes 3 (8B) through Ollama, 4-bit quantized (Q4_K_M GGUF) to fit in ~4.3–4.7 GB. Windows, Linux and macOS, with an 8 GB RAM floor. On Apple Silicon (M4), Metal acceleration gives 40–50+ tokens/sec.

OllamaQ4_K_M GGUF~4.5 GB8 GB RAM40–50 tok/s (M4)

Hosted · via OpenRouter

Call a model in the cloud

No local compute needed. OpenRouter gives API access to dozens of models, including free tiers — a good fallback when your machine isn't powerful enough.

OpenRouterno GPUfree tiers

$ ollama run qwen2.5:7b-instruct-q4_K_M
pulling manifest... done
$ ollama run hf.co/NousResearch/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M
>>> ready

$ export OPENROUTER_API_KEY=sk-...
$ agent.model = "openrouter/qwen-2.5-7b-instruct"
>>> connected

Watching the agent think

Every step is visible: what the agent decided, which tool it called, and what came back.

> user:  what's my Sui balance?
> agent: calling get_balance()
> tool:  12.4 SUI
> agent: You have 12.4 SUI.

Part 03

Agentic wallets on Sui.

The agent reasons in natural language, but every action it takes is a real, signed transaction on Sui.

01plain-language request 02agent reasons 03proposes a transaction 04you review 05you sign 06on-chain

What we're building

Agentic Wallet · Sui0x7f…a2

12.4 SUI

Ask your agent…

Send 2 SUI to alice.suiREVIEWED

The agent proposes the transaction in plain language; you stay the one who signs.

Sources

Everything referenced, in one place.

SuiThe Layer-1 the wallet signs tosui.io → Sui BasecampThe event this workshop was given atsui.io/basecamp → Sui documentationMove, transactions, the TypeScript SDKdocs.sui.io → OllamaRuns quantized models locallyollama.com → Qwen 2.5 (7B)The open-weights model used locallyhuggingface.co/Qwen → Hermes 3 (8B)The alternate local modelhuggingface.co/NousResearch → llama.cpp / GGUFThe quantization format (Q4_K_M) that shrinks the modelgithub.com/ggml-org/llama.cpp → OpenRouterOne API in front of many hosted modelsopenrouter.ai → ReAct (2022)The observe → think → act loop, written uparxiv.org/abs/2210.03629 → three.jsThe hero animation on this pagethreejs.org → QRCode.jsGenerates the code below, in your browsergithub.com/davidshimjs/qrcodejs →

Take it with you.

This code points to this page. Scan it to open the workshop on your phone — it's generated in your browser, from wherever this page is loaded.