Be agent provocateur: India's internet needs an AI makeover

A while back, I gave my AI agent my credit card, verification code and expiry date, and asked it to buy soap on Flipkart, in preparation for an agentic commerce hackathon I was participating in. An hour later, I allowed it to give up, after it got stuck in interfaces built to be attractive to humans, but near impossible for an agent to log in and buy something.

The internet is built around the assumption that human beings sit before a screen, search, read, click, fill forms, compare options and complete a transaction. That assumption is beginning to break.

At the hackathon that followed, I built something for Shopify purchases, only to find that Shopify only allows agents 'authorised' by them. Shopify and Flipkart, partners in Google's Universal Commerce Protocol (UCP), are both adversarial to open-source, independently-run agents in their own way.


Also Read: Who owns the bug when AI writes the code?

Many online businesses shun interoperability and lock users in, forcing agents to mimic human behaviour to bypass blocks. Amazon recently blocked Meta's Muse, the hottest agent out there. Many others block agents, while others ask you to solve a puzzle to prove that you're human.

Open-sourced agents like OpenClaw and Hermes have heralded the age of agents. While they're easy to set up, they are complicated to maintain and connect. Personal bots like Instinct, Muse, OpenAI's Dots and SpaceXAI's Grok remove much of this complexity, and are now adding users at unprecedented pace.

We're clearly moving from an app economy to an agentic economy. Apps are a means to an end, force us to fit their workflow, and have limited, siloed context about us. Agents do the opposite: they collect context from these silos via connectors, form a context graph of our preferences through interactions with us, and adapt to our needs.

At another hackathon, I imported my social media history of almost 20 yrs into my own memory vault, and allowed my agent access to my views, preferences and history. Data makes them easier and more intuitive for us to use. My agent generates my packing lists and suggests the right flights. Some day, a personal humanoid will pack for me.

Agent traffic on the internet is about 39% as according to Cloudflare Radar, and about 18% of India traffic. While we don't know how much of it is human-initiated, when agent traffic crosses 50%, it will be a moment of reckoning: will bots begin to matter more to online businesses than humans?

Given that the shift to an agentic web is now irreversible, here are 6 things India needs to get there, sooner the better:

Address adversarial environment Currently, the internet increases agents' navigation costs, and time it takes for them to execute actions. We need to adopt open protocols like Model Context Protocol (MCP) to allow websites and apps to expose resources, tools and structured inputs to AI applications.

Also Read: AI regulations coming soon in India: Govt to release consultation paper within a month, says Vaishnaw

Upgrade e-commerce catalogues A human can look at a product page and infer quality and reliability. An agent needs exact data and easier discovery. A minor change in a page layout, or an unexpected pop-up - meant for cross-selling and upselling to humans - can interrupt the entire process.

ONDC (Open Network for Digital Commerce), using the Beckn Protocol, has machine-readable catalogues and structured orders, and could be an agent-native commerce layer that predates MCP. But it's sadly underutilised.

Public services India's government websites are famously torturous for humans to navigate, let alone poor agents. India has built important DPI by separating the underlying rails from applications that use them. Government interfaces can expose authorised, intelligible and interoperable ways for an agent to act on a citizen's behalf: departments can publish safe-action APIs.

More intuitive mechanisms for payments One shouldn't have to give one's agent his or her credit card. While India has UPI AutoPay and e-mandates, we still await the launch of NPCI's Unified Agent Protocol. How this will play out in environments like insurance and lending remains to be seen.

Safe way to delegate authorisation in regulated environments Not every interaction warrants the same level of authentication. How we bring proportionality to identification of agents - because that involves attribution of that data to a human - needs to be seen from a privacy lens as well. The public internet offers a spectrum from anonymity to verification to both humans and their agents, and thresholds should rise with risk.

Clarity on liability How liability evolves with agents acting proportionately and autonomously is anybody's guess. But it needs clarity. Agents can make mistakes, and their behaviour is rarely deterministic, as they're impacted by compounding probability.

Their memory, the AI model they use, and the harness that controls their actions are variable, and no two outcomes are likely to be the same for two people, or arrived at in the same manner. We need to manage risk. But stack on liability, and people will stop using agents. Agent-insurance and easier reversibility of actions might be solutions.

Agents help us address key barriers to digital adoption: most of the internet is not Indian languages-friendly. Online interfaces are overwhelming and not intuitive for many users. Agents address that complexity by translating the web for us, bypassing the need to learn the web the way Western companies designed it for their users. Just like we in India leapfrogged wireline and adopted a wireless internet, we should be early adopters of an agentic web, for the benefit of our citizens.