Five years ago, I designed and implemented a blockchain network. The dream was to connect a large number of entities and create an ecosystem with a single source of truth, shared transactions, and none of the clunky enterprise integrations. The technology worked. But the world was not ready yet. The sponsoring organisation ended up running the nodes on behalf of the network participants as well, and exposing APIs to them. We had built a decentralised network, and then centralised it so that people could use it.

I have been thinking about that project while watching how agentic and AI-enabled solutions...

Writing code is becoming a lot like driving a car. By code, I do not mean programming as we understand it today, but the code that operates an agent.

Note that I did not say manufacturing a car, but driving one. A car takes you from point A to point B. The agents you build and operate do something similar. They help you solve a problem, sometimes over and over again. In some ways, tomorrow’s AI-enabled SaaS is like a taxi you hire. Someone else owns it, maintains it and drives it, and you pay for the journey.

Then there...

For the last few months I have been working on a project to classify requests from our store colleagues. The classification has 150+ classes. With a small set of classes it is relatively easy to get it right. Sentiment analysis, for example, is positive, negative or neutral. With 150 classes there are 150 ways things can go wrong, or even more.

I started with what seemed the obvious approach. I gave the model the user’s message and all classes with their descriptions in a single prompt, and asked it to pick one. Zero-shot classification. It was a gamble, and it...

A couple of weeks ago I worked with our business team on applying generative AI to a large dataset - around 130,000 rows in a spreadsheet. The model kept trying to retrieve and summarise. That appears to be its default behaviour. The team had tried everything reasonable. They split the data into smaller chunks. They wrote step-by-step instructions to walk the model through the retrieval. The answers kept drifting.

Then it struck us. The model was searching, not filtering or aggregating. Search is typically an approximate, probabilistic operation. What we needed was deterministic - the kind of result that comes...

The Mythical Man-Month by Frederick P. Brooks Jr. is a collection of essays on software engineering, first published in 1975. Though it is popularly known for its main theme of adding manpower to an already delayed project, the book is not about project management. It is about the balance between design and effort, between labour and thought.

In the age of AI agents writing code, we may rethink how the essays and concepts in the book apply in an organisation of agents. Here are a few essays with the original idea and a reframing in the context of agent organisation.

I often write in analogies… how the new AI wave compares to the dotcom boom or the industrial revolution. There is something to be learned from every transformation, whether it is from the factory floor or the financial markets. When we remove the specifics of the technology advancements, it becomes an economic phenomenon or a supply chain problem.

Looking at the way software is written by AI, I cannot help but compare it to 3D-printing in the manufacturing domain. 3D-printing was supposed to revolutionise manufacturing, but as of today it is still a rounding error in the overall manufacturing sector....

When we employ people for a job, we verify their credentials - their educational qualifications, certifications and work experiences. In the world of AI agents, similar to people, AI agents will need credentials. Human beings acquire theirs through the education system. We are trained for fifteen to twenty years, and that training qualifies us for certain jobs. The proof is in the certificates issued by universities and the experience letters from the companies where we worked.

AI is recent, and agents are more recent still. So how do we make sure they hold the credentials to perform critical tasks in...

Three decades, twelve editors. From vi in 1995 to Helix in 2026. The languages, jobs, and habits each one carried. A nostalgic look back, written before editors became viewers.

vi. (1995)

That’s how it started. We had Wipro Unix server with dumb terminals. Writing mostly C programs for hobby. C, Fortran

Turbo C Editor. (1998)

Not a real favourite. It had keyboard shortcuts and mouse support though. It took me through final year project to design concrete structures based on structural analysis outputs, constraints and assumptions. C

Tedit. (1999)

Terminal editor of Tandem computers. T6530...

There is a pattern in AI user interfaces, especially for the chat-based conversational interfaces. User asks a question or explains the problem. AI responds with a good-enough answer. User then follows up with more relevant information until an acceptable answer or solution is achieved. This suits the probabilistic nature of AI. This was the most common pattern in retrieval-augmented generation (RAG) architectures.

Contrast this with a human-to-human conversation where the first question will be responded with a counter-question or request for clarification, instead of a good-enough answer. As models became more capable and tokens became cheaper, we have introduced reasoning...

In 2011 I worked on Oracle Spend Classification, a data mining project. Machine learning algorithms classified products and parts into various spend categories for better control over the spend, contracts and even the design. Unlike the previous efforts the machine learning ran autonomously with a great degree of accuracy. But it worked on structured data, the product master, purchase orders, goods receipt and so on.

However, the world is mostly unstructured. The things we see, the things we do, the words we speak, the music we play. Computers are remarkably inefficient at making sense of all that. We have to...