Google Maps for AI Agents

Google Maps for codebases? That's basically what we built for our AI agents!

AI agents are on the rise, nowhere more so than in the software development industry. LLM's are language powerhouses that make them exceptionally adept at writing code in any programming language. To date they have been helpful sidekicks for quick prototypes (e.g. BOLT, Loveable, etc.) or code assistants (Cursor, Microsoft Copilot, etc.), specializing in small jobs or being helpful amplifiers to human engineers.

So no surprise that the most common projects we have been asked to support are AI agents capable of truly large assignments:

• agents that can rebuild a legacy system.
• agents that can perform data migrations.
• agents that can build a new CRM.

But modern software applications are BIG, and many of today's AI agents suffer two critical shortcomings:

• "Perspective" - being able to scale focus between macro and micro
• "Endurance" - being able to persist continuity and direction on very large initiatives

Humans have the innate ability to step back and see the big picture, or focus in on a particular detail, and to maintain our bearings through a large initiative.

What would it take for our AI agents to replicate these abilities?

I once asked a colleague for advice on adapting to my new role after a promotion. His answer was simple: "Andre, it's like in math: once you can add single digits, all you need to do to work in millions is move the decimal!"

And this captures the essence of the perspective problem. Being able to shrink things down, yet still retain the essence needed for the problem at hand. This insight led to the development of our "map" tool for AI agents. It works very much like google maps: zoom out and see the whole earth, or zoom in to a particular address. And with this, our agents gained the ability to choose their own focus.

I asked one of our AI coding agents: "Of all your tools, which is your favourite and why?" It picked the map tool, and had the following to add:

"In practice, I find this tool is often my starting point for almost any code-related task."

"Before diving into specific implementations, I use it to orient myself and build that crucial big picture understanding."

"It allows me to build a mental model of how different components relate to each other, rather than just seeing isolated files."

In a future post I will share more of my thoughts on solving the endurance problem, and how we are helping AI maintain continuity and direction on very large initiatives.