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

2026-06-16 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, quick

Summary

PreAct is a novel system designed to accelerate computer-using agents on repeated tasks by compiling successful runs into small state-machine programs. Unlike agents that re-reason every step, PreAct replays these compiled programs directly, achieving 8.5-13x faster execution without incurring per-step language-model calls. The system incorporates a crucial screen-state verification during replay, ensuring accuracy and reverting control to the agent if discrepancies arise. Furthermore, a rigorous store-time check validates that newly compiled programs genuinely solve tasks, preventing the accumulation of faulty programs and improving performance by 1.75-2.6 tasks per benchmark across mobile, desktop, and web platforms. PreAct also includes a fallback mechanism for novel situations, maintaining performance against strong record-and-replay baselines, and found that prompt wording, runtime guardrails, or the program selection method did not matter.

Key takeaway

For AI Engineers developing computer-using agents for repetitive workflows, PreAct demonstrates a critical path to significant efficiency gains. You should consider implementing a program compilation and replay mechanism to accelerate task execution by 8.5-13x, drastically reducing language model inference costs. Crucially, integrate robust screen-state validation and independent program evaluation to ensure reliability and prevent the accumulation of faulty automation, thereby maintaining consistent performance across diverse platforms.

Key insights

PreAct enables computer-using agents to learn and accelerate repeated tasks via compiled state-machine programs with robust validation.

Principles

• Agent efficiency improves by compiling successful task runs.
• Screen state validation is crucial for reliable program replay.
• Rigorous program validation prevents performance degradation.

Method

PreAct compiles a successful agent run into a state-machine program. This program is stored only after an independent evaluator confirms task completion. During replay, it checks screen state before acting, reverting to the agent if mismatches occur.

In practice

• Implement state-machine compilation for repetitive agent tasks.
• Integrate screen-state checks into automated UI interactions.
• Use independent evaluators to validate learned agent programs.

Topics

• PreAct
• Computer-Using Agents
• Task Automation
• Program Compilation
• State Machines
• Agent Efficiency

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Workflow-GYM: Towards Long-Horizon Evaluation of Computer-use Agentic tasks in Real-World Professional Fields — Current AI agents struggle with long-horizon, domain-specific GUI workflows, achieving only ~30% success.
• Build agents that run automatically — Automated agents can significantly enhance software development pipelines by handling review, monitoring, and routine tasks.
• New Stanford study reveals when teaming up AI agents is worth the compute — Multi-agent AI systems' performance advantage often stems from increased compute, not inherent teamwork superiority.
• The Truth About Agents in Production — Effective AI agent development prioritizes type safety, robust evaluation, and a problem-first approach over complex multi-agent systems.
• Long-Running AI Agent Browser Automation Tasks Is Here — AI agents demonstrate high persistence and adaptability in achieving complex, open-ended goals through browser automation.
• DeskCraft: Benchmarking Desktop Agents on Professional Workflows and Human-in-the-Loop Collaboration — DeskCraft benchmarks desktop agents on complex, human-collaborative professional workflows, revealing current AI limitations in long-horizon task execution.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.