Agentic perception loop — continuously perceive, reason, and act. The agent maintains observation memory and can use VLM reasoning to pursue goals.
Subcommands
Command
Description
agent run
Run the agentic loop locally on camera or video
agent start
Start the agentic loop on a running server
agent stop
Stop the agentic loop on a running server
agent status
Show agent status (running, tick count, plan, goal)
agent memory
Show recent observations from memory
agent recall QUERY
Query memory with natural language
Local Agent Loop
openeye agent run --model yolov8 --goal "monitor workspace safety"
openeye agent run --model yolov8 --video demo.mp4 --max-ticks 20
openeye agent run --vlm
Flag
Description
--model, -m <name>
Detection model (default: yolov8)
--goal, -g <text>
Goal for the agent to pursue (default: "monitor the scene")
--hz <float>
Loop frequency in Hz (default: 1.0)
--video <path>
Video file input (fallback if no camera)
--max-ticks <n>
Stop after N ticks (default: run until stopped)
--vlm
Enable VLM reasoning. Uses NEBIUS_MODEL (default Qwen/Qwen2.5-VL-72B-Instruct) with NEBIUS_API_KEY; a lowercase OpenRouter-style model ID uses OPENROUTER_API_KEY instead
Server Agent Control
These talk to a running openeye serve at OPENEYE_SERVER_URL (default http://localhost:8000) unless you pass --server, -s <url>.