Challenge 6: One-Shot a Multi-Agent Web App (Final Boss)

Global Hack Week: Agent Week | Backboard.io

Time needed: 20 minutes (write the prompt slowly, run it once)
Difficulty: Hard, in the best way. This is the flex challenge.
What you need: Challenges 1 to 5 complete. Your one-shot prompting muscles should be warm.

CHALLENGE WALKTHROUGH VIDEO HERE

The final challenge

One prompt. One shot. And this time the output is a real web app where multiple AI agents with different jobs work together to produce something none of them would produce alone.

Same one-shot rules as Challenge 5: write the prompt, fire it, hands off. If it misses, improve the prompt and re-fire from scratch. Answering your build agent’s clarifying questions is allowed. Patching with follow-up instructions is not.

What is multi-agent architecture?

Instead of one AI doing everything, you create a team of specialists:

Each agent has one job and instructions for exactly that job.
Agents hand work to each other: output of one becomes input of the next (and some can loop, like a reviewer sending work back).
A coordinator (your app’s code) runs the relay race.

Why do teams beat soloists? Focus. An agent whose entire personality is “you are a ruthless editor, find problems” catches things a do-everything agent glosses over. You are about to see this yourself.

How this maps to Backboard (this is why the API makes multi-agent easy):

Multi-agent concept	Backboard feature
One specialist agent	One assistant with its own system prompt (its job description)
One agent working one task	A thread on that assistant
Different brains for different jobs	Per-message model choice: 17,000+ models, pick per agent
Shared team knowledge	Memory and documents on an assistant, shared across its threads
An agent that acts	Tool calling and web search per message

Docs your build agent should use: (concepts + API reference).

Pick your app (or invent one)

Each of these has 3+ specialist agents and a web interface:

AI Newsroom: enter a topic → Researcher agent (web search) digs up facts → Writer agent drafts an article → Editor agent critiques and demands a revision → final article appears on the page with each agent’s contribution viewable.
Startup Simulator: enter an idea → CEO agent shapes the vision → Engineer agent proposes how to build it → Marketer agent writes the launch tweet → Investor agent grills the plan and scores it out of 10.
Story Studio: enter a premise → Plotter agent outlines → Author agent writes the scene → Critic agent reviews → Author revises → final story shown with the behind-the-scenes debate.
Debate Arena: enter any question → Pro agent argues for, Con agent argues against (2 rounds each) → Judge agent declares a winner with reasoning. Extra fun if Pro and Con run on different models.
Homework Helper Squad: enter a topic → Explainer agent teaches it → Quizmaster agent writes practice questions → Coach agent checks your answers and tracks weak spots with memory.

Requirements checklist (what “counts”)

Your app must have:

At least 3 assistants with genuinely different system prompts (different jobs, not three copies).
A real workflow between them: at least one agent consumes another agent’s output. A review-and-revise loop earns style points.
A web interface: input on a page, results on the page. A simple local Flask or FastAPI app is perfect. No need to host it online.
Visible teamwork: the page shows what each agent contributed (tabs, cards, or a timeline), not just the final answer. This is your demo money-shot.
Built in one shot on the Backboard API.

Example one-shot mega-prompt (AI Newsroom)

This is the level of detail a one-shot web app needs. Use the structure, but build your own idea.

Build a complete multi-agent web app called "AI Newsroom", then run it and
prove it works. One shot: build everything from this prompt alone.

Tech:
- Python backend with Flask, single-page frontend (HTML/CSS/JS served by
  Flask). Clean, modern dark theme.
- All AI via the Backboard API using the backboard-sdk pip package (docs:
  https://docs.backboard.io). API key from the BACKBOARD_API_KEY env var,
  never hardcoded.
- On first run, create THREE assistants and save their ids to agents.json
  for reuse on later runs:
  1. "Researcher": system prompt makes it a fact-finder. It must be sent
     messages with web_search="Auto" and must return 5-7 key facts with
     source links.
  2. "Writer": system prompt makes it a punchy news writer. It turns the
     Researcher's facts into a 300-500 word article with a headline.
  3. "Editor": system prompt makes it a tough but fair editor. It reviews
     the draft, lists 3 specific improvements, then the Writer revises
     once based on that feedback.

The workflow when a user submits a topic:
Step 1: POST /api/report kicks off Researcher (new thread) on the topic.
Step 2: Researcher's facts go to Writer (new thread) to draft the article.
Step 3: Draft goes to Editor (new thread) for critique.
Step 4: Critique goes back to the Writer's SAME thread for one revision.
Step 5: Respond with JSON: facts, draft, critique, and final article.

Frontend requirements:
- A topic input box and a "Run the Newsroom" button.
- A live status line while agents work ("Researcher is digging...",
  "Writer is drafting...", "Editor is grumbling...").
- Results in four tabs: Facts, First Draft, Editor's Notes, Final Article.
- Each tab labeled with which agent produced it.

Rules:
- Handle errors: if any agent call fails, show a readable error on the
  page, not a crash.
- requirements.txt and a README.md with run instructions.

Definition of done: start the server, run the full pipeline on the topic
"the rise of AI agents", confirm all four tabs fill with real content, and
tell me the exact command and URL to try it myself.

Notice the prompt decides everything in advance: agent count, personalities, thread strategy, API routes, UI tabs, error handling, and the proof-of-life demo. That is what one-shot readiness looks like.

Fire the shot

mkdir final-boss && cd final-boss
export BACKBOARD_API_KEY="your_key"

Start backboard (or open the folder in Backboard Studio).
Paste your mega-prompt. Press Enter. Watch a whole app assemble itself.
When it finishes, open the URL it gives you (usually http://localhost:5000) and run your app for real.

Re-roll rule: same as Challenge 5. Broken result means a smarter prompt, not follow-up patches. Two or three shots is normal for something this size. Each re-roll, diagnose which missing sentence caused the failure.

How to complete this challenge

Submit:

Your one-shot prompt.
A screenshot or (better) a short screen recording of the app in the browser, showing the multi-agent results.
Your code (GitHub repo link preferred). Make sure your API key is NOT in the code, that is what the environment variable was for.

Show it off in the Global Hack Week Discord too. You one-shotted a multi-agent web app. In one week you went from “sign up for an account” to this. That is worth bragging about.

Troubleshooting

Problem	Fix
App runs but agents sound identical	Your system prompts are too similar. Give each agent a strong, specific persona and forbid it from doing the others’ jobs.
Pipeline dies mid-run	Add the retry-and-show-errors rule to your prompt. Also check credits at app.backboard.io.
Frontend never updates	Tell the prompt to show errors on the page and log them to the console, then check the browser console in your re-roll.
It built something way simpler than you wanted	Every feature you did not write down is a feature you did not ask for. Spell out routes, tabs, and agent behaviors explicitly.
Port already in use	Close other running servers or ask for the app on a different port in your prompt.

You made it

Six challenges: account → tour → tools → agent → workflow → multi-agent app. You now know Backboard’s dashboard, Nash, R-CLI or Studio, and the API’s assistants, threads, memory, web search, and multi-agent patterns. Keep the docs bookmarked at https://docs.backboard.io, keep building, and come show the Backboard team what you make next.