NakodaX

NakodaX blog/Perspective·3 min read

Beyond AI Coding: Building AI Developers That Think With You

Why I moved away from AI SDLC frameworks and built my own personal AI developers to build NakodaX, and why decisions matter more than code.

Pallav Boonlia

By Pallav Boonlia

Founder & CTO, NakodaX ·

A thoughtful man with glasses and a greying beard rests his chin on his hand, reviewing a screen with a soft green glow in a dim office at dusk, with rows of quiet green lit monitors behind him.

Most developers are writing code blindly

In my opinion, most developers today are writing code blindly.

AI tools can generate code faster than anyone can review it. But speed is not the real problem. The real problem is that developers do not see the decisions the tools are making for them. The explanations are not well summarised, so the reasoning stays buried inside long conversations and agent logs.

I see this close to home. A developer I work with just writes code and commits it. Why was this approach chosen? What else was considered? What will it cost us later? Nobody asks.

Even AI SDLC frameworks did not help me here. They gave me a process, but they did not help me understand what was actually being built.

What I changed while building NakodaX

I moved away from AI SDLC and switched to Dots and Duet. With them I built a set of personal virtual developers who work for me and build NakodaX.

The idea was simple. I wanted agents that work independently, understand the product, keep context, write code, review changes and help monitor ongoing development.

But I had one important requirement. I did not want to become just another person approving whatever the AI generated. I wanted the AI to understand how I make decisions.

The big problem was decision making

The big problem was how to improve decision making the way I do it myself.

When I work on architecture or product decisions, I do not simply pick an option. I think about the business requirement, technical constraints, security, scalability, maintainability and the impact on the existing system.

This is where Jev and Gemini decision models became stronger for me. I applied a combination of both as a separate decision layer, so decisions become structured, explicit and reviewable before the work moves forward. It asks questions like:

  • Should we introduce a new service or extend an existing one?
  • Is this implementation aligned with the existing architecture?
  • Does the proposed change add unnecessary complexity?
  • Are there security or performance implications?
  • Should the agent proceed, reconsider or ask for approval?

The result is that building code and product is stronger than ever for me. It does not guarantee that every decision is correct, which is exactly why human oversight still matters.

Diagram titled Dots build, Jev checks, and I approve what matters. Dots agents send decisions to Jev, which checks architecture, complexity and security. Checked decisions go to an approval box with the proposal, reasons, alternatives and risks, then to the architect, who approves the important decisions. Dots learn the architect's decisions and style. Duet and Gemini feed in with dashed lines, and Claude and Cursor send decisions summarised back to the approval box.
how Dots, Jev and I connect · 4 steps and a learning loopDots agents send their decisions through Jev, the checked decisions reach me in a box for approval, and what I decide teaches the Dots agents. Claude and Cursor send their decisions to the same box. Gemini and Duet are still new to me, so they are drawn dashed.

My Dots agents are learning how I decide

One great thing about Dots is that the agents are learning my decisions and my style. Over time they pick up the standards I expect and which decisions matter to me. They have started treating me as their friendly neighbourhood architect.

Duet is still new to me, so I do not have much feedback on it yet. I will share more once I have used it properly.

Decisions come back to me for approval

I still use Claude and Cursor for development. The difference is that decisions are summarised back to me in a box for approval. In that box I see:

  • What is being proposed
  • Why this approach is recommended
  • What alternatives were considered
  • The risks and implications
  • What needs my approval

The summary reaches me by email. The email links to a dashboard where I approve or change each call, and my decision is recorded back in Dots.

This gives me visibility without having to supervise every line of code. The goal is not to slow development down. It is to keep important architectural and product decisions visible and accountable.

A message in an email inbox from NakodaX headed 2 decisions need your approval. Two cards show a data room closing date banner and a Chrome extension session change, each with what is proposed, the risk, three check scores and a Review button.
The email
A decisions dashboard in a light theme with a dark sidebar. The queue lists two decisions that need approval and one that went ahead. The Chrome extension decision is open, showing what is proposed, why, the risk, check scores and Approve and Reconsider buttons.
The dashboard

Where I think this is going

I believe we are moving towards multiple specialised agents working together instead of one assistant doing everything. Some will build, others will review, test, monitor and evaluate decisions. Humans stay in charge of direction, judgment and accountability.

For me it comes down to four things:

  • Context. Agents understand the product, architecture, history and my preferences.
  • Decision intelligence. Decisions are structured and made visible, not hidden inside conversations.
  • Autonomous execution. Personal AI developers write, review, test and monitor software.
  • Human oversight. Important decisions come back to the architect or product owner for approval.

This is not about removing developers. It is about giving developers more leverage while keeping engineering judgment at the centre.

What comes next

I intend to open source this work, including the workflows and the decision process, once I have cleaned up the remaining work on NakodaX. I would love to see how other teams use it.

One thought I keep coming back to. We have spent years teaching AI to write code. Perhaps it is time to teach our AI developers how to think about the decisions behind that code.

What is your experience with AI driven development? Are you confident in the decisions your coding agents make, or are you still reviewing code without enough visibility into why it was written that way?

NakodaX | Building with intelligence. Protecting what matters.

See which product fits what you share.

Start a free trial, or tell us what you share and with whom, and we will show you how NakodaX handles it.