ASPLOS '26 Tutorial: Building Agentic AI Systems with Jac

Learn to build agentic AI systems using Jac — a language designed around the 7 primitives of agentic computation. By the end of this hands-on session you will have built a self-correcting, parallel research agent from scratch, one primitive at a time.

---

Repository Structure

exercises/          <- You work here (skeletons with TODOs)
solutions/          <- Reference solutions (peek if stuck)
output/             <- Generated markdown files (created at runtime)

---

What You Will Build

Mind primitives — what the LLM does:

Step	Primitive	What it does
1	Generate	LLM returns free text from a function signature
2	Extract	LLM returns structured, typed data — enforced by the compiler
3	Invoke	LLM calls tools, observes results, and loops (ReAct cycle)

Flow primitives — how work moves:

Step	Primitive	What it does
4	Pipe	Chain operations sequentially
5	Route	LLM picks a path from a graph of nodes
6	Loop	Repeat until a typed quality check passes
7	Spawn	Run multiple walkers in parallel and merge results

Step 8 puts all 7 together into a single self-correcting parallel research agent.

---

Setup

1. Run the setup script

source setup.sh

This will:

• Install the Jac language runtime (standalone binary)
• Add jac to your PATH

2. Set your API key

You will receive an OpenAI API key at the start of the tutorial. Export it in your terminal:

export OPENAI_API_KEY="your-key-here"

---

How to Work Through the Exercises

Each exercise file has TODO markers where you write the key primitive. The boilerplate and types are provided.

1. Open the exercise

# Work in the exercises/ directory
cd exercises/

2. Fill in the TODOs

Open step1.jac in your editor. Look for the # TODO comments — they tell you exactly what to write.

3. Run your code

jac run step1.jac

4. View the output

Each step writes a markdown file to output/. Open it to see your results formatted nicely:

# On macOS
open output/step1_generate.md

# Or just cat it
cat output/step1_generate.md

5. If you get stuck

The complete solution is in solutions/:

jac run ../solutions/step1_generate.jac

---

Exercise Progression

# Mind primitives
jac run step1.jac    # Generate: LLM returns free text
jac run step2.jac    # Extract:  LLM returns typed data
jac run step3.jac    # Invoke:   LLM calls tools (ReAct)

# Flow primitives
jac run step4.jac    # Pipe:     Sequential chaining
jac run step5.jac    # Route:    LLM-directed graph traversal
jac run step6.jac    # Loop:     Self-correction cycle
jac run step7.jac    # Spawn:    Parallel walkers

# Capstone
jac run step8.jac    # All 7 primitives in one agent

---

Resources

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