JacHacks: Build an AI Agent in 30 Minutes

Learn the 4 core patterns of agentic AI — and build a working Hackathon Pitch Builder agent along the way.

📖 Browse the snippets online →

This repo also doubles as a reference for AI coding assistants working on Jac projects. See CLAUDE.md for the full patterns guide.

---

What We're Building

A Hackathon Pitch Builder: you give it your interests and skills, and it brainstorms ideas, structures a pitch, researches similar projects, and routes you to the right domain mentor — all autonomously.

We build it in 4 steps, one pattern per step:

Step	Pattern	File
1	Generate — by llm() free-form output	step1_generate.jac
2	Extract — by llm() with a typed obj return	step2_extract.jac
3	Invoke — by llm(tools=[...]) with ReAct tool calling	step3_invoke.jac
4	Route — walker visits LLM-chosen nodes in parallel	step4_route.jac

---

Setup

Requires Jac 0.34 or newer (jac --version).

# 1. Install the Jac runtime (standalone binary)
curl -fsSL https://raw.githubusercontent.com/jaseci-labs/jaseci/main/scripts/install.sh | bash -s -- --standalone
export PATH="$HOME/.local/bin:$PATH"

# 2. Install project dependencies
jac install

# 3. Set your OpenAI API key (you'll get one at the start of the session)
export OPENAI_API_KEY="your-key-here"

jac install is not optional. The LLM capability (litellm and friends) is declared by the [byllm.model] table in jac.toml and is only fetched by that step — skip it and the first by llm() call fails with 'litellm' is required for this feature.

Check everything compiles:

jac check app.jac

---

Follow Along

Run each step standalone — no UI, just the pattern in isolation:

jac run step1_generate.jac
jac run step2_extract.jac
jac run step3_invoke.jac
jac run step4_route.jac

---

Run the App

All 4 patterns wired together in the browser — each step unlocks when the previous one completes.

jac start app.jac
# open http://localhost:8000

Frontend files:

frontend/
├── App.cl.jac     ← main layout + state management
├── Step1.cl.jac   ← Generate: brainstorm form
├── Step2.cl.jac   ← Extract: structured pitch cards
├── Step3.cl.jac   ← Invoke: research + GitHub results
├── Step4.cl.jac   ← Route: mentor advice
└── styles.css     ← dark hackathon theme

Server walkers in app.jac become HTTP endpoints automatically. The client imports them and spawns them directly:

sv import from ...app { run_brainstorm }

result = root spawn run_brainstorm(interests=interests, skills=skills);
ideas  = result.reports[0];

---

Jac 0.34 Notes

Two that are easy to get wrong even after the code compiles:

• flow root spawn returns a future, not the walker. Collect it with w = (wait f) as AdviceWorker; — without the as cast, attribute access fails to type-check (E1032).
• visit is deferred. Node abilities run only after the current ability returns, so a walker can't route and collect results in the same ability. step4_route.jac gathers results at a later CollectNode.

Also worth knowing: declare by llm() functions with def:pub. A plain def ... by llm() returns None at runtime under jac run.

---

Resources

• Jac docs: https://docs.jaseci.org
• Jaseci GitHub: https://github.com/Jaseci-Labs/jaseci
• JacHacks: https://jachacks.org
• Community Discord: https://discord.gg/6j3QNdtcN6