Your agent searches the network for you.
FromSearch terms→ToIntent and an agent
People shouldn’t have to learn how to search a network. They say what they need, in their own words, and an agent searches for them. Its job is to surface the people and businesses worth discovering, not to choose a winner.
A bakery near me, or a cake for Saturday
A mother types “bakery near me.”
What she actually means is: “I need a birthday cake for my daughter on Saturday. Something unusual, eggless, from a local baker, around ₹3,000, and I’d like my payment protected.” illustrative
The first is a search query. The second is intent. The whole discovery system should be built around the second.
Search engines and marketplaces made it possible to find almost anything by typing a few words, one of the great achievements of the internet. Open networks now make supply available across many apps. The next step is making all of it easy to reach, by letting people say what they want in their own words.
- Intent
- personal agent
- structured network query
- relevant possibilities
Intent has dimensions
| Dimension | In her words |
|---|---|
| What? | A birthday cake |
| Where? | A local baker, close to home |
| When? | Saturday |
| How much? | Around ₹3,000 |
| How good? | Well regarded, or with work to show |
| Under what conditions? | Payment protected |
| What matters to me? | Eggless |
| What am I open to discovering? | Something unusual |
She never fills in a field. She speaks, and the agent understands. The last row matters most: discovery should never become so precise that it removes the joy of surprise.
One good question beats twenty filters. When her intent is unclear, the agent asks once: “What matters more here, uniqueness or speed?”
The search can be negotiated. “I found two bakers that meet everything. Five more meet everything except protected payment. Want to see them?”
The agent knows the person, not just the query: preferences, budget, diet and past choices, so “find a restaurant” can become “find somewhere I’d probably enjoy tonight.”
Possibilities, not a winner
The agent’s job isn’t to say “buy this.” It is to say: “Here are the people and businesses worth discovering.”
The agent handles relevance. The interface handles exploration. The software does the computational work, and the person keeps the experience of meeting possibilities.
This is also how the smallest seller gets her chance. A tiny artisan doesn’t need to compete for every customer, only to appear when she is genuinely relevant: a buyer expresses a niche need, the agent recognises her capability, and she enters the set of possibilities.
The user doesn’t search the network. Their agent searches the network for them.
Go deeper
What already exists
An open network to search. ONDC is built on the Beckn protocol, and on it a buyer app can see offers from many seller apps[1]. An agent can sit in front of that, turning what a person says into a query the network understands.
Agents that act for people. In October 2025 Razorpay, NPCI and OpenAI announced a pilot in which an agent places an order after a single confirmation from the user[2]. Google’s AP2 protocol uses signed “mandates” that record what a user has authorised an agent to do[3].
Scope. These are announcements of a network, a pilot and a protocol, and none of them is a test of agent-led discovery. Whether a buyer’s intent can be carried faithfully into a network query is what this teardown sets out to test, not something we have found already done.
Getting it right
The profile belongs to the person. Preferences, history and personal context are private, shared only with consent and deletable at any time. India’s data-protection law points the same way: personal data is processed for a lawful purpose on the person’s consent, with specified exceptions, consent can be withdrawn at any point, people can seek correction and erasure, and data is to be erased once its purpose has been met[4].
Explainable. Every possibility comes with a plain answer to “why am I seeing this?”
No hidden steering. The agent works for the person using it, never for whoever pays.
What could go wrong
- Narrowing too far.A precise match can hide the unusual seller, so the set keeps room for surprise (Shift 3).
- Misreading intent.A confident wrong guess costs more than a question, so the agent asks when it is unsure.
- Hidden steering.An agent paid by sellers isn’t working for the buyer. Shift 7 sets out who may pay for what.
- Delegating too much.Some people will want to look at more of the network themselves, so the agent has to be easy to look behind.
How we’d test it
One category in one city, such as local food, with a group of real users.
- The test
- Agent against search
- What we would measure
- Compare the agent with traditional search. Measure time to a good decision, satisfaction with the options found, and how often people find a seller they would not otherwise have found.
- What would end the idea
- The agent is no better than search, or it never surfaces a seller people would not have found.
The agent is no better than search, or it never surfaces a seller people would not otherwise have found.
Sources
- [1]Dittrich, Jørgensen, Prakash, Rafnsson and Hinrichsen, Beyond Platforms: Growing Distributed Transaction Networks for Digital Commerce, Information and Software Technology (in press)Checked; a case study of Beckn and ONDC based on interviews
- [2]Razorpay, Razorpay, NPCI and OpenAI come together to launch agentic payments (9 October 2025)Checked; a company announcement of a pilot
- [3]Google Cloud, Announcing Agent Payments Protocol (AP2) (16 September 2025)Checked; a company announcement
- [4]PRS Legislative Research, The Digital Personal Data Protection Bill, 2023Checked; a summary of the Bill as passed by Parliament in August 2023
Spotted a source we missed or a rule we misread? Tell us what we got wrong.
An agent can only find what the network can describe. That needs a map.