Recipe Add Multiple Recipes
Add the following recipes into the Broccoli app: Recipe: Sweet Potato and Black Bean Tacos description: A quick and easy meal, perfect for busy weekdays. servings: 3-4 servings preparationTime: 4 hrs ingredients: as desired directions: Roast sweet potato cubes, mix with black beans, and use as filling for tacos. Top with avocado and cilantro lime sauce. Try adding a pinch of your favorite spices for extra flavor. Recipe: Salmon with Dill Sauce description: A quick and easy meal, perfect for busy weekdays. servings: 3-4 servings preparationTime: 10 mins ingredients: per individual taste directions: Bake salmon fillets and serve with a sauce made from Greek yogurt, dill, lemon juice, and garlic. Try adding a pinch of your favorite spices for extra flavor. Recipe: Caprese Salad Skewers description: A quick and easy meal, perfect for busy weekdays. servings: 8 servings preparationTime: 1 hrs ingredients: as needed directions: Thread cherry tomatoes, basil leaves, and mozzarella balls onto skewers. Drizzle with balsamic glaze. Garnish with fresh herbs for a more vibrant taste.
- Task outcome
- Pass
- Pass on both graders1
- Time per task
- 186.2 s
- vs. Mobilerun 429.7 s22.3× faster
- AI model cost
- $0.0537
- vs. Mobilerun ~$3.683~68× cheaper
Recording and steps
No memory vs “shared memory”
Mobilerun used “shared memory” carried lessons from task to task. Lusor agent ran with clean context and no memory of prior task runs or attempts.
No self-improvement loop vs “self-improved”
Mobilerun’s self-improvement agent reviewed its own benchmark runs and rewrote its app cards and memory before each evaluation round. Lusor’s agent doesn’t learn between runs: our team wrote its six app skills and general guidelines by hand, and all of them are published.
Select skills vs “app cards”
Mobilerun used per-app instruction files (“app cards”); which apps had them, and what they said, wasn’t published. Lusor created small skills for 6 of the 21 AndroidWorld apps. The agent had to decide to use them. They are fully published.
Run details
- Recorded
- November 9, 2026
- Phone
- Android 13 emulator, reset to the benchmark snapshot
- Lusor tool calls
- 45
1Grader: Mobilerun’s fork of the AndroidWorld grader (c5467e6); Google’s unmodified grader (30583e2) also shown. 2Speed = Mobilerun’s published session time ÷ ours. 3Mobilerun est. $ = their published action count × $0.0766 (one GPT-6 Astra low call per action at the per-call cost we measured); their token use is not published, so it is an estimate, not their bill. “Skill assisted” = the agent read an app skill that has rules for that task (research’s reading of each skill’s text); reading a skill that does not apply is not marked.