Exploring AI trading for the Indian market.

A powerful framework, built for developers.

TradingAgents is a serious piece of open-source research. Ten AI specialists examine a stock from different angles and argue their way to a recommendation, the way a trading desk would.

It also assumed a developer: someone who already knew which stock to look at, was comfortable at the command line, could configure the project, and could read the output.

It was also built around US markets, and it forgot everything between runs. So the question I started with was simple: what would this look like if it began with the trader’s day rather than the engine?

The design opportunity sat right at the front. Help someone work out what deserves attention, before asking the AI to go deep on it.

What’s worth looking at today?

The original started with a ticker. This starts with the market. Open it and you see where things stand before you have typed anything.

Nothing to search for. The morning view loads before you ask it a question.

Three decisions shaped this screen. Start with context, so the first thing you read is what the market is doing. Surface things worth investigating, already computed, rather than waiting for a query. And give an obvious next step, repeated in the same place on every screen, so nobody wonders what to do next.

Which stocks deserve a closer look?

Before running anything expensive, the system scans the market and puts forward a shortlist. You can browse as long as you like, expand any suggestion to see why it was surfaced, and only then decide whether it is worth a deeper look.

Looking is free. The AI only runs on something you have already decided is worth it.

That separation is the point. If every glance feels like it costs something, people stop glancing, and a tool you are afraid to browse is a tool you stop opening.

A scanner tuned to the kinds of movement that matter on NSE, rather than a generic one.

Why is this one worth considering?

Now the AI runs. Ten specialists look at the same stock from different perspectives, two of them argue the case for and against, a risk review pushes back, and a decision comes out the other end with entry and exit levels attached.

The design idea underneath: the AI should not feel like a black box. You can watch each step as it happens, read both sides of the argument next to each other, and see how the recommendation was arrived at rather than just being handed it.

Showing the reasoning, not just the answer.

Did the idea actually work?

Any recommendation can be tracked as a simulated trade instead of a real one. It records what the idea was and where it came from, then shows what actually happened to it over the following days and weeks.

Outcomes over time, grouped by which screen produced the idea. Losses tracked the same way as wins.

The principle is that trust should come from seeing what happened, not from believing the recommendation. A few weeks of tracked ideas turns “should I believe this?” into something you can look at.

The .env file, killed.

To use the original, you had to find a configuration file, paste a key into it, and run the project hoping you had got it right. For anyone who is not a developer, that is where the exploration ends.

Test tells you the key works before you spend anything running an analysis.

A small interface decision, and the one that removed the biggest barrier between the system and anyone wanting to try it.

A simple loop.

The interesting part was not any single screen. Once they were connected, and each one carried its context into the next, the thing started behaving like a loop: an idea could be tracked, the outcome could be reviewed, and what came back could shape what got looked at next.

Four observations from the exploration.

• AI is a feature, not the product. The workflow around it mattered more than the model behind it.
• Make complex systems understandable. Showing the process was more useful than hiding it behind a single answer.
• Context changes the experience. Market hours, NSE and BSE, and how people here actually trade all change what the product should do.
• Small interactions decide adoption. The settings screen mattered because it removed the barrier to everything else.