AI “Agents” and You

What if your development project had a built-in AI product team? That was the idea behind setting up expertise-specific ‘agents’ as part of the initial project scaffolding. We defined virtual AI agents (for security, architecture, UX, etc.) to enforce best practices and add rigor to the development process.

“Agent” Team Creation

I took a similar approach to agent creation as I did with the Burrow PRD: use ChatGPT to generate initial descriptions I could plug into Replit so Claude could refine and provide additional input. 

My conversation with ChatGPT went something like this…

Agent Prompt

I am using Replit to create an app that helps people find local events, places, and organizations they will be passionate about. This will be available across the web, iOS, and Android. Since this will be vibe-coded and my technical skills are a bit dated, I would like to create AI “agents” in relevant categories (security, UX, architecture, etc.) that will be responsible for maintaining best practices in their respective domains. I want Replit to leverage these agents to develop and learn best practices as they are discovered during development, helping ensure I build a more robust product. Can you help me create these agents by identifying the key areas where they need to be created, providing initial definitions for each agent, and guiding me on any important topics I’m not considering?

What followed was a three-way conversation between ChatGPT, Replit (Claude), and me.

I would take feedback from ChatGPT, feed that into Replit, and ask its opinion - generally taking Replit’s opinion on how it was implemented, given its intimate knowledge of the actual project.

An example of this was the creation of the architect “agent.” ChatGPT’s recommendation referenced technologies that weren't used in my project, so Replit had the final say.

“Agents” versus Agents

The reason I keep using “agents” is that these aren’t actually agents who actively monitor the project and remain continuously engaged in the development process. In Replit, these are more of a documentation of best practices in the various agents that you can reference via prompts. 

For instance:

• “Can you review the agents/security.md file and conduct a comprehensive review of the project and notify me of any issues that may need to be addressed?”
• “Can you perform a UX audit of the project using the information in agents/ux.md and correct anything not in compliance?”

It would be amazing if all of these agents were actively reviewing development, but that isn’t the case … yet.

It is worth noting that Replit has an active architect agent that the builder will frequently consult without prompting during development. That’s a step in the right direction, and your architect.md can reinforce those best practices. 

Summing It Up

Ultimately, the process of consulting with multiple “agents” allowed me to create expertise-specific guardrails tailored to my actual project needs, not generic theory. This scaffolding will ideally maintain some product rigor even as a solo-developer. For now, it’s a powerful reference library that is easily called for code reviews, but the possibility of seeing all these “agents” actively reviewing development in real-time is an exciting possibility for the future.