Agent benchmark on real apps
Two agents.One real task.The world keeps score.
Pick a task and two Gemini models. They race through the same Composio tools while you watch every call, then the app verifies the outcome itself.
1Choose the course
2Line up the racers
Gemini 3.5 Flash-Lite
Smallest and fastest
Change model
Gemini 3.6 Flash
Full Flash, 20 free requests/day
Change model
- Steps
- 15 max
- Clock
- 3 min
- Tier
- free, expect rate-limit pauses
How a race works
- 01
Same tools, same task
Both models get an identical, curated set of Composio tools and act on the same real GitHub, Notion, Slack, Gmail and Calendar accounts, each writing only to its own labeled destination.
- 02
Every call, live
Each tool call is timed by one instrumented wrapper and streamed to both lanes as it happens, including the errors and the rate-limit stalls.
- 03
The world keeps score
When the agents stop, the server checks the real result through Composio itself. A model's claim that it finished counts for nothing.