A party-planning agent that books vendors through their own booking websites (test sites in this demo), counts a booking only when the vendor confirms it, survives crashes without double-booking, and keeps every decision in an audit log in MongoDB Atlas. Lyfe Events plans a whole celebration from one chat message, typed or spoken in English, Spanish or French. The concierge (Google Gemini with function calling) asks for any missing details, then proposes a plan from 157 Toronto vendors that fits the budget, food always included, with trend ideas and real "Near you" shops for favours and decorations. When the host confirms, a LangGraph booking agent fills in each vendor's own website form. A step counts as done only when the vendor's booking page shows the reference number: "sent" is not "booked". Spend over $1,000 waits for the host's approval, in the chat or through an Auth0 by Okta login. If the agent crashes, it resumes from the verified state in MongoDB Atlas and asks each vendor what it already has, so nothing is booked twice. When a vendor changes its form, Atlas Vector Search (Voyage embeddings) finds a similar form the agent handled before and the agent adapts. Every plan, policy check, approval, form read, send, verify and retry is written to a schema-validated `audit_log` collection, and the host can stop the run from the page (a change stream on `controls`). The app also designs invitations (wording by Gemini, checked by IBM Granite), imports guests from a CSV, tracks RSVPs, shows the event-day forecast (Environment Canada data, The Weather Network link) and speaks briefings with ElevenLabs. Live demo: https://lyfe-events.tail93d927.ts.net:8443/ Honest limits: Vendor booking sites are test sites on our own server; other catalogue vendors get a "contact them directly" card. Invitations go to a simulated outbox. The demo server runs on the team's laptop behind a Tailscale https link. The idea comes from the Lyfe Events business plan; all code was written during the event. See DISCLOSURE.md.
More from Agents in Context Toronto: a MongoDB Hackathon
We simulate real world agent to agent systems built on MongoDB, Langchain (harnesses) ready for use in production to defend all systems. A swarm of real AI…

suture is an incident-recovery agent that finishes the job even when its worker crashes mid-run. Cleaning up a leaked key takes a chain of writes across…

Venues play one playlist for everyone, and the apps that let guests shape the music keep their likes on view with no easy way to take them back. Musory: your…

Robots are starting to learn from people, and fleets can share what one robot learns with every other robot. Nobody can yet answer three questions: who taught…

Phylax Tower is a safety net for airport control towers. In March 2026 at LaGuardia, a jet was cleared to land, and two minutes later a fire truck was cleared…
New hires hit the ground at light speed. New devs lose their first days juggling broken setup guides, each solving the same problems alone. Day One finishes…
