Gifting Agent

Inspiration

We sell into B2B revenue teams and live in the corporate gifting world: ~$250B a year moves through it, and most of it lands in the bin. Every deal gets the same branded notebook, because choosing a gift a recipient actually likes is hard — and generic LLMs can't help. They'll happily write a lovely note about a gift they invented, but they have no idea what a "RevOps leader at B2B SaaS" audience actually loves. The hackathon's premise — agents can reason but are culturally blind — described our exact pain, so the idea came fast: give a gifting agent real cultural grounding with Qloo.

What it does

Gifting Agent turns deal context into a gift + note, through a five-step loop that is fully visible on screen:

1. PLAN — strategy from deal state alone. A $250 close-won celebration plans differently than a $120 post-outage apology.
2. TOOL — three live Qloo calls: segment affinity tags, seed-entity resolution, then cross-domain taste neighbours of the seed.
3. SELECT — candidates must fit the budget AND the gift history; a domain already sent is skipped.
4. CHECK — self-checks each candidate out loud (budget / freshness / evidence) and re-plans visibly on failure.
5. COMPOSE — the note, in the occasion's tone, opening on shared cultural ground.

And when every domain its budget reaches is already used, it refuses to send and escalates to a human rather than repeat a gift — a repeat weakens the relationship, and the agent knows it.

How we built it

Cloudflare Worker (free tier): static demo UI served as an asset, /api/* as the agent loop in TypeScript calling Qloo's REST API directly with X-Api-Key — never in the browser, never committed. Affinity data describes groups, not individuals: input is segment-level, and anything shaped like an email or phone number is rejected before any call. Every pick cites the exact Qloo evidence behind it; if Qloo is unreachable, bundled fixtures keep the demo running and the UI honestly labels the run DEMO DATA. Thread state is browser-owned (localStorage) and sent with each request, so the server answers deterministically from what it is given.

Challenges we ran into

• Workers isolates don't share memory. Our first version kept thread state server-side and counters drifted between requests — one request saw "1 touch," the next saw 3. We made the server stateless: the client owns state, the server is pure.
• Qloo's untyped /v2/insights returns zero entities. We probed every entity type individually and found filter.type is required, and place returns museums and attractions — which turned out to be a gift for gifting (experiences!).
• Parallel bursts to Qloo flake. Six concurrent insight calls intermittently failed mid-test. We switched to per-domain calls with graceful degradation: a domain without an entity neighbour falls back to citing the segment's affinity tags instead of inventing.
• Every API-shaped error taught us the real response shapes. The fix loop was: observe the actual JSON, unwrap, redeploy.

Accomplishments that we're proud of

• Fully live and hosted: judges can run the whole loop in about two minutes with no login.
• The agent's reasoning is legible — the trace shows plan -> Qloo tool calls -> selection -> self-check -> composition, so you can audit why a gift was chosen.
• Guardrails that act: budget ceilings, history-aware selection, and a visible refusal-to-repeat escalation.
• Privacy by design, not as an afterthought.

What we learned

• Qloo's taste graph is genuinely cross-domain — resolving one seed into experiences, music, film, brands and books is what makes a gifting agent possible at all.
• On Workers, design for statelessness first; memory between requests is a trap.
• Agent quality is mostly guardrail quality: knowing when not to send is the feature.

What's next for Gifting Agent

• An optional LLM layer that composes free-form note variations grounded on the cited Qloo evidence, with the same trace around it.
• A feedback loop: record whether the recipient loved it, and let gift-domain strategy learn per segment.
• Portfolio mode: plan gifts across an entire deal portfolio with per-deal budgets and a shared "taste diversity" constraint across accounts.
• CRM/webhook integrations so deal stages arrive automatically instead of being picked in the demo.

Built With

• agentic-ai
• cloudflare-workers
• css
• html
• localstorage
• qloo
• qloo-api
• rest-api
• taste-ai
• typescript
• wrangler