Give your agent a reason to look again.
We built AutoVision for the OpenCV AI Competition 2026, powered by AWS, to explore a question:
What if a vision system could decide when it needs a closer look?
OpenCV 5 measures the image. A Bedrock-powered agent evaluates the evidence, chooses a follow-up tool, and uses the actual result to decide what to do next.
Perceive → Decide → Inspect → Re-evaluate → Recommend human review
More than a single detection
- Inspect a region: crop a detected object and run detection again.
- Improve a dark frame: apply low-light enhancement and remeasure.
- Ask a focused question: request an additional scene assessment.
- Recover from rejected tools: return the error to the agent so it can choose another approach.
Watch the workflow happen
The working console shows live perception, Decider choices, tool execution, and results. Saved reports let you inspect the evidence and replay the recorded steps.
We also published evaluation results with the failures intact. A completed workflow does not guarantee a correct conclusion, and AI recommendations remain subject to human review.
Bring vision to your own application
AutoVision is a standalone service with a web console, REST API, and MCP tools for compatible agents. CloudFormation templates support private deployment into your own AWS account.
Try it: https://autovision.dev
Explore the code: https://github.com/techcto/autovision
Follow the judges’ walkthrough: https://github.com/techcto/autovision/blob/main/docs/JUDGES_FIELD_GUIDE.md
What should an agent look at twice?
We would love to hear from other builders working on computer vision, agent tools, and evidence-driven workflows.
Where would a second look make the biggest difference in your application?
Log in or sign up for Devpost to join the conversation.