Inspiration
What it does
Inspiration
Haggling is where money leaks for most people. Sellers expect negotiation on used goods, but most buyers either skip it (and overpay) or dread it (and give up). Meanwhile, AI agents are starting to transact on our behalf — but handing an agent a wallet without guardrails is scary. We wanted to show what trustworthy agentic commerce looks like: an agent that negotiates hard, can never break your budget, and still asks you before a single cent moves.
What it does
- Pick a used item (camera, bike, headphones, console, watch) and set your maximum budget, plus optional requirements in plain language ("must include the lens cap and a working battery").
- Your buyer agent negotiates with the seller's agent for up to 6 rounds. Every offer is shown in a live chat-style log with the agent's reasoning, and a guardrail bar shows the current offer against your max budget.
- The buyer agent can NEVER agree above your budget. If the seller's floor is out of reach, the agent walks away — and explains the final gap.
- If they agree, you see the deal summary and savings, approve it, and the payment is settled with PayPal (sandbox Orders API: create → approve → capture).
- The result screen leads with what matters: "You saved $XX (YY% off)". The full audit log then reveals the seller's secret floor price, so you can verify the negotiation was real.
A one-click Judge Mode runs both a winning deal and a walk-away with zero typing.
How we built it
Next.js (App Router, TypeScript) as a single app with in-memory state. The negotiation engine is deliberately hybrid: offer amounts come from a deterministic concession strategy (buyer target, seller personality — stubborn / balanced / eager — and round pressure), while an LLM, when a key is provided, only phrases the messages. Money math never depends on a model. PayPal is integrated via the REST Orders API against the sandbox, with a clearly labeled mock mode with identical flow so judges can run everything with zero credentials. A natural-language parser extracts budgets and requirements from free-text notes.
Challenges we ran into
- Making agent negotiation feel real without letting an LLM decide prices: solved by separating strategy (deterministic) from speech (optional LLM).
- Trust: an agent that can overspend is a non-starter, so the budget cap is enforced in the strategy layer and at the payment layer — a payment request above the cap is refused even if a deal state claims otherwise.
- Judges' time: the whole story (win + walk-away) had to fit in under 3 minutes, which led to Judge Mode presets.
Accomplishments that we're proud of
A complete, honest negotiation loop: deals under budget, walk-aways when the math doesn't work, the seller's hidden floor revealed afterwards for verification, and a real
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for HagglePal
Built With
- next.js
- paypal
- typescript
