AI-assisted software development makes building new features to help our customers an exercise in speed.
Rainforest has a deep culture of experimentation and iteration, and we’re actively exploring how AI can help us work smarter. At our core, we aren’t just building an AI test generation tool; we are constantly experimenting with how AI can make our own development cycles leaner, faster, and more intuitive. That includes experimenting with AI-assisted software development.
Recently, we hit an architectural crossroads while evolving our core product. We’re moving away from “one-shot” AI test generation to a conversational, step-by-step UI. Here’s what that means and why it matters to us and our customers (plus takeaways for other devs and software teams).
The Challenge: Context is King
Until now, our engine worked in a linear fashion: Yyou gave it a natural language prompt, and it returned a full end-to-end test.
It was powerful but lacked the nuance of human collaboration, like the back-and-forth that naturally happens when working with a team (and ultimately results in a better product). So we wanted to give our users the ability to chat with the AI, generating a few steps, course-correcting, and then building more.
The technical hurdle? Memory. For a chat-like experience to feel “human,” the AI agents need a persistent memory of the conversation history and the previously generated steps.
The Brainstorm: Capturing the Lightning
Last week, I sat down with our backend team to map out the design. We needed an architecture that could handle stateful conversations without bloating our infrastructure.
Once upon a time, turning a complex design problem like agent memory into a functional, well-tested implementation would have required at least a full day of documenting and coding. Today, AI-assisted software development makes it possible to let the machines do much of this work.
So I decided to let the machines do the heavy lifting from minute one.
Step 1: The Raw Input
First, I captured the team meeting transcript (using Krisp, for those curious, which also happens to provide great noise cancelling). Once we hopped off the call, I used its AI summarization to distill 60 minutes of technical debate into a structured architectural map.
Step 2: The First Draft
I fed that summary into Antigravity using the Gemini 3 Pro model. Within a few minutes, it laid out a functional backbone. Unfortunately, it also took a few shortcuts that didn’t sit right with me: It gave some classes too many responsibilities, creating a tight coupling between components that would have made the system difficult to maintain or evolve as it grows.
This is the “‘human-in-the-loop”‘ moment, where an engineer’s knowledge is needed to steer the AI toward a cleaner, more modular architecture.
Step 3: The Refinement
This is where the magic happened. I took the original AI summary, added a few specific architectural guidelines (defining exactly where those responsibilities should live), and opened Cursor, using the Opus 4.5 model to try out some AI-assisted software generation.
The Result: 40 Minutes to “Ready”
The output was, frankly, impressive. It delivered a solution that met every requirement we discussed in the meeting. Because the context was so well-preserved from the transcript summary to the final LLM, the code needed only minor adjustments.
Less than 40 minutes after the meeting ended, we had a production-ready design and the core logic implemented.
Why AI-Assisted Software Generation Matters
We share this anecdote because we believe the “how” of building software is just as important as the “what.” By optimizing our internal resources and using the same type of agentic workflows we build for our customers, we can move faster and focus on what really matters: making software testing easier.
We’re excited to roll out this new chat-based test generation feature soon. It’s built on a foundation of solid architecture with a little bit of AI-assisted speed.
How is your team using AI-assisted software development? What tools are you loving? Hit us up on LinkedIn or X and tag us when you share your projects.
P.S. Want to get early access to this feature? We’d love to show you how the platform works and what we’re building next.
