PreAct: Computer-Using Agents that Get Faster on Repeated Tasks

Paper: arxiv.org/abs/2606.17929 · Website: 01.me/research/PreAct · Author: Bojie Li (Pine AI)

What this is

Computer-using agents drive real software through the screen — clicking and typing — but they solve every task from scratch: asked to repeat a task, an agent re-reads the screen, re-reasons every tap, and pays the full cost again.

PreAct lets such an agent get faster on tasks it has done before. The first time it succeeds, PreAct compiles the run into a small state-machine program — states that check the screen, transitions that act — and on later runs replays it directly instead of invoking the agent: 8.5–13× faster, with no per-step language-model calls.

Two checks make this safe:

• Replay-time check. At each step PreAct verifies that the screen matches what the program expects before acting, and hands control back to the agent the moment something is off.
• Store-time check (verify-before-store). A freshly compiled program enters the library only if, re-run from a clean state, an independent evaluator confirms it actually solved the task — catching programs that replay to their last step yet leave the task undone.

Across a mobile (AndroidWorld), a desktop (OSWorld), and a web (WebArena) benchmark, the store-time check is what separates repeated runs that improve from ones that degrade as faulty programs accumulate — worth 1.75–2.6 tasks per benchmark, the same direction on all three. A fallback that explores afresh when no program fits brings PreAct level with a strong record-and-replay baseline. The paper also reports what did not matter: prompt wording, runtime guardrails, and whether a language model or a plain embedding retriever selects which program to reuse.

Repository layout

Path	Contents
preact/	The PreAct library: compiler, executor, RAG program store, LLM clients, per-platform agents
benchmark/androidworld/	AndroidWorld runner (run_docker.py)
benchmark/osworld/	OSWorld runner (run_osworld.py)
benchmark/webarena/	WebArena runner (run_webarena.py) + environment setup helpers
paper/	arXiv preprint source (latex/preact.tex) and findings tables
site/	Project website (React + Vite) — source for 01.me/research/PreAct
RESULTS.md, EVALUATION_REPORT.md	Full empirical results and per-experiment write-ups

Setup

PreAct itself is a Python package; the three benchmarks each wrap an upstream environment that you install separately.

# Python 3.11+
pip install -e .

Set the API keys for the models you intend to use:

Variable	Needed for
ANTHROPIC_API_KEY	Compilation + program selection (always), and the Claude computer-use agent on OSWorld / WebArena and optionally AndroidWorld
GEMINI_API_KEY (or GOOGLE_API_KEY)	The Gemini 3 Flash agent used by default for AndroidWorld steps

Benchmark environments (each is the upstream project; follow its own install guide):

• AndroidWorld — a Docker image exposing the HTTP control server on port 5000: sudo docker run --privileged -p 5000:5000 -it android_world:latest
• OSWorld — a local ~/OSWorld checkout providing evaluation_examples/<task-set>.json and a VM/Docker backend.
• WebArena — the shopping_admin Docker environment; helpers live in benchmark/webarena/setup.py.

Reproducing the experiments

The central result is the verify-before-store ablation, run as a cold→warm pair (empty library → reused library) with the gate on vs off. The gate is controlled per platform as noted below; all numbers in the paper come from these runners.

AndroidWorld — gate is the --verify-before-store / --no-verify-before-store flag (default on). The 15 task types of the official-15 subset:

TASKS="ContactsAddContact ContactsNewContactDraft AudioRecorderRecordAudio \
AudioRecorderRecordAudioWithFileName BrowserDraw BrowserMaze CameraTakePhoto \
CameraTakeVideo ClockStopWatchPausedVerify ClockStopWatchRunning FilesDeleteFile \
MarkorCreateFolder MarkorCreateNote SystemBrightnessMax SystemWifiTurnOn"

# gate ON (default) — the agent uses Gemini 3 Flash for steps by default
python -m benchmark.androidworld.run_docker --tasks $TASKS --n-instances 1 --seed 42

# gate OFF (ablation)
python -m benchmark.androidworld.run_docker --tasks $TASKS --n-instances 1 --seed 42 \
  --no-verify-before-store

OSWorld — gate is the same flag (default on); task set is read from ~/OSWorld/evaluation_examples/<task-set>.json:

python -m benchmark.osworld.run_osworld --task-set test_tiny                          # gate ON
python -m benchmark.osworld.run_osworld --task-set test_tiny --no-verify-before-store # gate OFF

WebArena — gate is the PREACT_VERIFY_BEFORE_STORE environment variable (on default / off):

PREACT_VERIFY_BEFORE_STORE=on  python -m benchmark.webarena.run_webarena   # gate ON
PREACT_VERIFY_BEFORE_STORE=off python -m benchmark.webarena.run_webarena   # gate OFF

Other ablations (environment variables)

Each remaining experiment in the paper is selected by one environment variable:

Variable	Default	Values	Controls
PREACT_VERIFY_BEFORE_STORE	on	on / off	Store-time verify gate (WebArena; Android/OSWorld use the CLI flag above)
PREACT_CACHE_MISS_FALLBACK	skip	skip / cua	On a cache miss, skip the repeat run or explore afresh with the full agent (WebArena)
PREACT_SELECTOR_MODE	agentic	agentic / embedding	Pick the reuse candidate with an LLM tool call or a plain embedding retriever
PREACT_RUNTIME_MODE	state_machine	state_machine / flat_script	Replay as a verified state machine or a flat action script
PREACT_GUARDRAILS	on	on / off	Runtime code-level guardrails (AndroidWorld)
PREACT_CUA_PROVIDER	gemini	gemini / anthropic	Backend for AndroidWorld agent steps
PREACT_COMPILE_PROVIDER	(Claude)	gemini	Swap the compile-step LLM to Gemini for the cross-provider check
PREACT_MODEL	claude-sonnet-4-6	model id	Model used for compilation and program selection

See RESULTS.md for the per-experiment result tables and the exact runs behind each paper figure.

Citation

@article{li2026preact,
  title   = {{PreAct}: Computer-Using Agents that Get Faster on Repeated Tasks},
  author  = {Li, Bojie},
  journal = {arXiv preprint arXiv:2606.17929},
  year    = {2026},
}