Inspiration

Most AI agents constantly “think” even when nothing meaningful has changed. That wastes compute, adds latency, and makes behavior unpredictable. We wanted agents to react to reality, not timers.

What it does

Seconds Count adds a perception gateway in front of an AI agent. It watches the screen for change, analyzes cheaply with pixel diffs and OCR, and only authorizes an agent when uncertainty remains.

How we built it

We built a lightweight control loop that polls screen changes, runs OCR to detect loading or known states, and escalates to an agent only when cheap signals can’t explain what’s happening. The demo visualizes this with live state indicators.

Challenges we ran into

Creating reliable, fast signals from pixels and OCR without false positives was tricky. We also had to design a demo that showed value without letting an agent run wild.

Accomplishments that we're proud of

We built a clear, event-driven alternative to blind waiting that reduces unnecessary model calls and makes agent behavior understandable and debuggable in real time.

What we learned

Most agent problems aren’t intelligence problems—they’re perception and control problems. Cheap, deterministic checks dramatically improve reliability and cost.

What's next for Seconds Count

We plan to plug this gateway into real agents, expand perception signals, and use it to safely grant autonomy only when the environment truly demands it.

Built With

Share this project:

Updates