AuditGate is a customer support and refund agent built around a simple idea: an AI shouldn't be trusted to police itself with money. Instead of relying on system prompts to keep the model in line, AuditGate puts every financial threshold, security check, and approval rule directly inside MongoDB Atlas. When a customer asks for a refund—either by typing or speaking through ElevenLabs voice—the agent verifies their order in MongoDB and checks our live policy rules without ever revealing internal limits to the user. Routine refunds within the limit go through automatically. If someone asks for a larger amount, LangGraph pauses the conversation state in MongoDB until a manager reviews the request and hits Approve, which immediately wakes the agent back up via MongoDB Change Streams to finish the refund. To protect against jailbreaks, Atlas Vector Search screens incoming messages for prompt-injection patterns and pulls up relevant policy precedents for human reviewers. Every decision is permanently recorded in a schema-validated MongoDB audit log linked directly to its LangSmith trace, giving teams a clear, tamper-proof record of every action the agent takes.
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…
