Exchangent

Built for the YC Make something agents want hackathon with Surya Pugazhenthi.

We built Exchangent to allow agents to do the entire lifecycle of bartering and negotiating a deal. Most tools are for GTM like Clay, which handles active outreach, and Shopify have their own agents that can let people ask questions about a shop.

However from here nothing truly exists to have an agent actually handle a sale end to end without a human in the loop. So we built the infrastructure and system to do it.

Architecture

Beginning a Sale

The entry point is our public facing MCP or Connector. These integrate directly in Claude Code, Codex, or any harness that supports MCP/Connectors.

A user starts by just prompting their agent through any interface that's setup, even through just claude.ai or openai.com. Prompts ideally include the budget and maximum loss that's acceptable.

Find me a desk no wider than 48 inches and an ergonomic chair delivered in San Francisco in 5 days. Budget $400 all-in, I'd love $300.

Then the user's agent will use our MCP/Connector to place the buy order. These are hosted on manufact, and act as the gateway to begin an order process. This MCP then hits our backend to create a buy order transaction and kick off the buyer agent. We also do a match based on details of your order and available buyers. For each match we begin an A2A session using Inkbox.

Inkbox assigns an identity that a separate long running claude session uses to negotiate with a connected party, using its respective identity for its role.

Negotiating Agent Protocol

Then the haggling begins, we have one buyer agent haggle with n matches of agents.

Each buyer and seller agent also has a set limit of how much each party will "lose" in a deal, and this is governed by our system not agents. Of course all of this is set through our agent connector.

Each send over offers, which are saved as transactions in our database to ensure no offer is lost, along with reasoning about the price.

This is important for two reasons.

1. The great cure for hallucinations is determinism, and a database row is the absolute source of truth.
2. Yet sometimes there are reasons for prices, such as high quality, speed, timing. These are all factors needed for determining the value of an item and is "service".

The buyer agent can then agree to any price.

We call this NAP, Negotiating Agent Protocol, which is the infrastructure our long running agents use to barter.

Completing the sale

Agents will negotiate until either the buyer's lowest price is met, or a standstill occurs. The user is also able to ask the agent to accept an offer early.

Once done, the agent makes an offer transaction to the buyer, and once the buyer accepts the offer gets made into a Stripe checkout.

From here, just open the link, pay, and the sale is completed just through your agent.

Testing + Success

From our runs, sale periods often lasted at most 2 minutes before agents came to an agreement. As most people know, this is significantly faster than any human sale.

The experiment worked best with multiple sellers, and we had really cool UI (that my partner built) that showcased the process.

We were able to consistently do an entire flow from buy to end with a Stripe link sent and a purchase made.

Hurdles

By far the biggest issue was the vertical. We aimed to be a marketplace, which everyone knows is almost impossible to pitch as actually useful. Buyers and Sellers need to exist, imbalance creates major problems.

What was a better approach would be to instead be an integration layer. Where agents shopping the web can find your site, see a buyer agent ready for them, tell it what the user wants, then the buyer agent begins the bartering with all buyers integrated in.

Another issue was making the buyer/seller agent, as we needed tokens and we had none.

Luckily YC Startup School gave me a Claude Max plan, but without api tokens it would be against the TOS to have a VM run my Claude plan. And I didn't feel I should spend money on a hackathon. But low key I definitely should have.

But also it may have not mattered, as the main issue is just the marketplace itself.