Agentic Patterns
Atriva is designed to work alongside LLMs and AI agents, not replace them. The key insight behind the hybrid edge-AI architecture is that continuous cloud inference is impractical: sending every camera frame to a VLM costs too much, introduces unacceptable latency, and creates privacy exposure.
The correct architecture is:
[Camera] → [Atriva Edge Pipeline] → [Structured Event] → [LLM / VLM / Agent]Atriva handles the high-frequency, low-latency perception layer. Agents handle the low-frequency, high-reasoning layer. Together they form a cost-effective, privacy-first system that can autonomously observe, investigate, and act.
Core Concepts
Events as Agent Tool Inputs
Every detection Atriva makes can be emitted as a typed JSON event — a structured signal that an agent can receive as a tool call result or webhook payload. An event carries:
type— what was detected (e.g.,fall,ppe_violation,crowd_surge)confidence— model confidence scoretimestamp— ISO 8601 UTCcamera_id— source camerazone— configured zone within the camera viewsnapshot_url— link to the captured frame (optional, on-device or secure storage)metadata— detection-specific fields (e.g.,{ "missing_ppe": ["helmet", "vest"] })
This event schema is the handoff point between Atriva and your agent.
The Pre-Filter Principle
Atriva acts as a pre-filter: it discards the 99% of frames where nothing meaningful happens and surfaces only the events that warrant higher-level reasoning. This means your LLM or VLM is invoked tens or hundreds of times per day, not millions — keeping inference costs manageable.
Agent Action Patterns
Once an event reaches an agent, common action patterns include:
| Pattern | Description |
|---|---|
| VLM investigation | Pass snapshot to GPT-4V / Claude claude-sonnet-4-6 for deeper visual analysis |
| Report generation | LLM writes a structured incident report in natural language |
| Alert routing | Agent determines who to notify and through which channel |
| Evidence packaging | Agent collects snapshot, event metadata, and camera context into a record |
| Escalation decision | Agent decides whether to escalate based on severity and history |
Patterns in This Section
- Event-Triggered VLM — Call a VLM when an Atriva event fires
- Natural Language Alerts — LLM-generated alert messages vs. static templates
- Multi-Agent Smart Spaces — Per-camera agents with a supervisor agent
Compatible Frameworks
Atriva events work with any LLM or agent framework that can receive a webhook or call a REST API:
- LangGraph — model events as graph nodes that trigger tool calls
- CrewAI — assign a safety agent task when a violation event fires
- AutoGen — multi-agent conversation triggered by event data
- Custom agents — consume the event webhook with any HTTP client