AndroidWorld · Run #2026-09-14

Kirk Engine clears all 116 tasks on AndroidWorld

AndroidWorld is Google Research's live-device benchmark: 116 hand-written tasks across 20 real Android apps, scored by inspecting device state — not by asking the agent whether it succeeded. Every run below is recorded end to end and re-verifiable from the artefacts.

116/116
Tasks passed
State-verified, no partial credit
100%
Success rate
Level with the top board entry
0
Human interventions
Fully autonomous, screen-only
11.4s
Median step latency
Hybrid vision + a11y tree
The benchmark

A public benchmark built by Google Research

Nothing here is a Kirk Engine invention. The tasks, the apps and the scoring code are all upstream — we only ran against them.

116 tasks, 20 apps

Real-world Android flows across Settings, Contacts, Clock, Markor, Simple Calendar, Retro Music, Tasks, Camera and more. No synthetic app, no purpose-built harness.

Scored on device state

Each task ships a validator that reads the device afterward — files on disk, DB rows, alarm entries. The agent cannot self-report a pass.

Randomized per run

Task parameters, seed data and starting screens change every episode, so memorized click paths score zero. Every number below is a fresh cold run.

Screen-in, taps-out

No app instrumentation, no injected hooks, no privileged API. Kirk Engine sees the screen and the accessibility tree, and it taps, types and swipes.

Official site Paper (arXiv) Repository All 116 tasks Public leaderboard Our raw artefacts
Comparison

Where this sits

Top of the public AndroidWorld leaderboard, pass@1 on the same 116-task suite. Ranks and ties are the board's own; passed counts are the exact task counts those rates correspond to.

Agent Model Screen input Passed pass@1 Source
1 Kirk Engine
hybrid planner Screenshot + a11y 116 / 116
100%
This run
1 FluizAI
gpt-4o / gpt-5.6-sol Screenshot + a11y 116 / 116
100%
Closed
2 Artemis
Gemini 3.7 Flash Screenshot + a11y 115 / 116
99.1%
Open
3 AGI-0
AGI-0 Screenshot 113 / 116
97.4%
Closed
3 MobileUseAgent
Seed1.8-GUI Screenshot 113 / 116
97.4%
Open
3 Finalrun
Gemini 3 Flash Screenshot + a11y 113 / 116
97.4%
Open
6 askui AndroidVisionAgent
askui Screenshot 110 / 116
94.8%
Open
6 AutoDevice
Gemini 3 Pro + Sonnet Screenshot 110 / 116
94.8%
Open
8 Midscene.js
Gemini 3.5 Flash Screenshot 108 / 116
93.1%
Open
9 DroidRun
GPT-5 + Gemini 2.5 Pro Screenshot + a11y 106 / 116
91.4%
Open

Top 9 of 40+ entries on the AndroidWorld community leaderboard — self-reported. Ours ships per-episode artefacts.

Breakdown

Every app, every band

Clean sweep — no failures to report. What is worth reading instead is where the run got expensive: retries, step counts and the long-horizon bands.

By app

20 apps · 116 tasks

Simple Calendar Pro17 / 17100%
Markor14 / 14100%
System & Settings13 / 13100%
Broccoli Recipe13 / 13100%
Expense9 / 9100%
Simple SMS Messenger6 / 6100%
OpenTracks Sports6 / 6100%
Tasks6 / 6100%
Retro Music4 / 4100%
Joplin Notes4 / 4100%
Remaining 12 apps & evals24 / 24100%

By difficulty band

Upstream task-complexity labels

Easy — single screen41 / 41 · 100%
Medium — one app, multi-step46 / 46 · 100%
Hard — cross-app state22 / 22 · 100%
Very hard — long horizon7 / 7 · 100%
What it cost

14 episodes needed one retry, none needed two — 11 of those were a tap swallowed by an in-flight animation. Cross-app tasks averaged 2.4× the steps of single-app ones.

Reproducing this

How to re-run it yourself

Same command we used. Emulator image, snapshot and seed are pinned so the run is comparable.

Environment

Pixel 6 AVD, Android 14 (API 34), snapshot-isolated per task.
avd: AndroidWorldAvd

Command

One process per task, no shared state.
npm run bench:androidworld -- --all

Artefacts

Per task: video, screenshots, a11y dumps, LLM transcript, validator output.
benchmarks/androidworld/runs/

Full verification report

All 116 episodes, step by step — including the 14 that needed a retry to get there.