agent architect

Inspiration

Starting an AI agent project is difficult when you have an idea but little architecture experience. Agent Architect helps beginners clarify what to build before jumping into code

What it does

Agent Architect asks up to five clarification questions and requires the user to confirm a requirement summary. It then produces an architecture blueprint, Python starter code, and a Critic review with source-checked quotations. Results can be downloaded. It completes an initial design workflow, not the implementation of a finished business application.

How we built it

Built with Python, Streamlit, and the Strands Agents SDK, using DeepSeek through an OpenAI-compatible model adapter.

One Strands Agent is reused across interview, design, and review stages. Three tools retrieve official Strands documentation, local architecture patterns, and code templates. Application code enforces the question limit, confirmation gates, and output validation.

Challenges we ran into

Key challenges included validating model outputs, preserving user requirements across stages, and keeping Critic feedback grounded in actual source material. Intermittent generation failures and overly critical review suggestions remain limitations.

Accomplishments that we're proud of

We connected the full idea-to-design workflow, implemented real tool calls, and passed 26 automated tests. Three consecutive live-model flows also passed for one study-planning scenario.

What we learned

Clear permission boundaries and user confirmation matter as much as generated code. Checking that a quotation exists does not prove that the reviewer's judgment is correct.

What's next for agent architect

Test with beginner developers, improve generation reliability and review quality, and make starter code easier to extend. Business integrations still require implementation, and external user testing is pending.

Built With

• deepseek
• preserving-user-requirements-across-stages
• python
• strands
• streamlit