WHAT I'M WORKING ON
An agent-first protocol (rs/1) for Simulator Navigation
As you might know from previous newsletter issues: I'm working on a bundled RocketSim CLI. I want to reinvent how agents navigate through Simulators, and reduce token costs & time spent. RocketSim has the unique advantage of running in the background, allowing for optimizations other popular tools can't do.
I created 5 testing scenarios using the system settings app installed on every Simulator. Comparing the latest RocketSim CLI version to popular open-source alternatives, the results are quite impressive:
- ~12x fewer tokens
- 1.6x faster execution
- 4x fewer wrong taps
RocketSim is able to do this due to its unique agent-first protocol (called rs/1) instead of relying on raw AX dumps. It also has a live snapshot store which allows the agent to reuse state instead of paying to rediscover the UI on every turn. Combined with semantic actions, waits, and typed errors, the agent models have to guess less and get faster to their end destination.
I've also spent a lot of time improving the discoverability of elements on screen. In my experience, elements like tab bars and navigation bars weren't always discovered by other tools.
Finally, installation will be easy due to a bundled CLI & Agent Skill, that will be symlinked from the settings screen commands:
I'm working hard this week and I hope to get it out of the door before next week, so stay tuned!