AWS Builder Center

# Weekend Showcase Challenge: IdeaStress

An autonomous, multi-agent "Second Brain" powered by AWS Bedrock Nova, Step Functions, and Weaviate that forces you to confront the reality of your startup ideas through rigorous AI stress-testing.

The Vision: From Ambiguity Paralysis to Execution
Every one of us has watched a friend—or been the friend—who announces a massive plan with total confidence: "I'm going to ship an MVP in two weeks and grab 10% of the market." The plan sounds great in the room because nobody wants to kill the vibe by pushing back.
But the gap between an unexamined assumption and a failed project is exactly where most startup energy gets wasted. I wanted to build "the friend" who does push back.
IdeaStress is an autonomous, multi-agent "Second Brain" that forces you to confront the reality of your ideas. It takes a vague concept and runs it through a gauntlet of specialized AI agents. An idea doesn't get to become a plan until it survives being questioned.

How it Works

You enter a free-text idea into the React app. From there, it enters a five-agent pipeline with two mandatory human checkpoints:

  1. Excavator: Reads the idea and surfaces hidden assumptions (timeline, market size, skills).
  1. Researcher: Checks each assumption against a Weaviate vector database seeded with startup failure data, falling back to live web search if needed.

  1. [Human Gate 1] The pipeline pauses. You review the evidence-backed assumptions.
  1. Adversary: Builds the strongest possible case against your plan, finding specific risks and the hardest question you haven't answered.

  1. [Human Gate 2] The pipeline pauses again. You read the attack and choose your pivot.

  1. Planner & Critic: Generates a 30/60/90-day plan with a literal "do this tomorrow" step. The Critic grades the plan, forcing internal revisions until the feasibility score is high enough.

For this Weekend Showcase, I took the build further by implementing an Export Plan to PDF feature, allowing users to take their dynamically generated, highly-scored execution plan offline to share with their co-founders instantly.

AWS Services & Architecture Overview

This isn't just a Python script; it is a highly scalable, serverless pipeline.

  • Amazon Bedrock (Nova Lite): Used as the reasoning engine for all five specialized agents.
  • AWS Step Functions: The orchestration backbone. We used the  waitForTaskToken  pattern to physically pause the execution at Gate 1 and Gate 2 until the human makes a decision.
  • AWS Lambda: Each of the 5 agents runs in its own isolated Python 3.12 Lambda, allowing us to deploy and update them independently.
  • Amazon API Gateway & AWS Amplify: Exposing the Step Functions to a beautifully designed, Glassmorphism-styled React/Vite frontend hosted on Amplify.

How I Built It (Process & Challenges)

The biggest challenge was handling the asynchronous Human-in-the-Loop (HITL) gates.  waitForTaskToken  is built for this, but two separate Lambdas (the one triggering the wait and the one the frontend polls) can't easily share state across cold starts.
I originally tried an in-memory dictionary, but it failed when the container recycled. To solve this, I engineered our Weaviate Cloud (WCD) vector database to also act as a persistent key-value state cache. By saving the  TaskTokens  to Weaviate, the Step Function can pause indefinitely while the human thinks, and the React frontend can poll the database to know exactly when to render the next UI screen.

What I Learned Across the Summer

I learned that the gap between a system architecture document and the deployed product is where the real engineering happens. Over the summer, I had to pivot my database schema (collapsing collections to fit sandbox limits) and my Bedrock models (adapting to quota limits dynamically).
More importantly, I learned that "Responsible AI" is easiest to defend when it's built as a hard constraint. By engineering genuine Step Functions pauses for Human Gates 1 and 2, there is no code path where the AI can silently skip past the human. The human always retains absolute agency.

Links & Shoutouts

  • Builder Inspiration: A huge shoutout to Lewis Sawe! Your "Museum That Grows" creative agent from the last challenge was a massive inspiration for how to build autonomous, always-on reasoning loops without losing the magic of the output.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article