At the AI Summit, learning to love and fear the era of agents

How does a reporter cover a massive international summit with simultaneous programming? That was my challenge when I entered the India AI Impact Summit at Bharat Mandapam in New Delhi last week.

I first thought I had an infallible tech-enabled plan: I would use a snippet of code to fetch audio clips of YouTube livestreams of several panel discussions, transcribe them with a cloud-based voice recognition tool, review those transcripts later, and then file stories.

But that is not what happened. The organisers seem to have found it difficult to get an adequate number of video and livestream personnel to reliably take care of a dozen simultaneous livestreams. My technique ran into trouble at several discussions, where the stream started only after the panel was introduced. This made it challenging for me to “diarize” transcripts (add speaker names to each text) by cross-referencing the panellists listed on the summit website. The website itself kept changing without notice. And sometimes, much to my annoyance, someone would leave a live mic on a production counter, severely compromising my automated transcription.

Thankfully, other reporting assignments kept me busy — daily briefings, interviews with different players in the AI ecosystem, and high-profile controversies, including the haphazard security arrangements that caused speakers to miss their own sessions and Galgotias University’s Chinese robodog.

As I ran around taking notes, I kept thinking, ‘there has got to be an easier way to do this’. And there indeed was, for many tasks. I felt guilty whenever I was not trying to find those easier ways. That feeling came from having just downloaded Claude, a coding-focused AI app that could help make my work much simpler.

With no coding experience, I built three tools likely to save me time on grunt work. The first is an Android app to continuously fetch notifications from the Gazette of India, which has a tendency to drop, with little warning, major policy updates through the week. The second is a process to update my personal website. While it earlier took me half an hour every three months to do this, it now takes just a minute. The third was a small browser extension to automate the many steps required to submit a drafted story.

I was — and still am — excited and scared. Excited because I can be free of some of the drudgery that is incidental to my work. Fear because these are capabilities that theoretically everyone has now. And with time, these may not be just cool hacks, but a core part of the capabilities we are expected to exercise. How can someone who grew up in a pre-AI internet era compete with someone who tries all they can to get more done?

To be sure, the craft of journalism is not as exposed to AI-led displacement as, say, programming. The profession demands that a journalist earn the trust of experts — and convince them to share what they know as well. The LLM frenzy is, however, likely to diffuse in unsettling ways. What happens to journalists starting out without these tools when their competition is using agents to compose and send individualised emails seeking information? What happens if the AI natives supercharge their productivity and compound their skills faster than someone who still does things as they have been for the last decade?

For now, there is likely a ceiling: complex workflows involving autonomous agents — as opposed to mere productivity hacks — require tokens for LLM inference, which are expensive. Even with a paid Claude subscription, I hit usage limits using its most advanced model to build the three tools above. These costs offer some cushion. But as inference becomes cheaper, professionals will have to face the full possibilities of automation and the costs of failing to adapt.