Research map / Multi-agent systems and games / LLM agents

Agent JIT Compilation for Latency-Optimizing Web Agent Planning and Scheduling

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Compiles natural-language web tasks into cost-optimized executable code (JIT-Planner) and selects serial/parallel/hedged execution via Monte Carlo latency estimation (JIT-Scheduler), with an invariant-enforcing tool protocol.

Problem

Computer-use agents built on per-step LLM loops and primitive click/type actions are slow and error-prone, and prior work ignores the space of cost-varying code plans and execution schedules.

Approach

Treat agent tasks as JIT-compiled code over cached reusable tools, with pre/postcondition tool manifests enabling static validation, CFG-based cost estimation to pick the cheapest valid plan, and Monte Carlo latency sampling from learned element-latency distributions to choose serial, parallel, or hedged execution.

Key results

• JIT-Planner achieves 10.4x mean latency speedup (11.7s vs 122.1s) and +28% accuracy over Browser-Use across 37 tasks.
• JIT-Scheduler with Gemini-2.5-Pro achieves 109.9s at 86.4% accuracy versus OpenAI CUA's 258.7s at 77.8%, a 2.4x speedup and +9% accuracy.
• Protocol enforcement raises GPT-4.1 valid-plan rate from 78% to 91% and Gemini-2.5-Pro from 79% to 96%.

Surprising findings

• Parallel hedging with 8 workers reaches a 95% valid-plan threshold faster than serial retry, even for GPT-5 (16s vs 23s).
• High task cardinality does not imply parallelization is fastest; some high-cardinality tasks are optimal under serial execution.

Limitations

• Requires one-time offline setup per application (tool synthesis and trace collection), best for repeatedly-used apps.
• Cached tools break under UI changes; relies on fallback re-planning.
• Evaluated only on web apps from REAL and WebArena (37 tasks); spurious failures in stochastic environments.

Experimental setup

Datasets: REAL benchmark tasks (Dashdish, Gomail, Omnizon) WebArena tasks (GitLab, Reddit) offline execution traces for caches

Benchmarks: REAL WebArena

Metrics: end-to-end latency (s) task accuracy (%) valid-plan rate Pass@k Pass@t speedup

Baselines: Browser-Use Browser-Use +cache OpenAI CUA Anthropic CUA CUA +cache Fixed-Scheduler (Serial, Parallel, Hedge) Oracle-Scheduler

Builds on: Browser-Use REAL benchmark WebArena Model Context Protocol

Closest papers by meaning

• The Tool Decathlon: Benchmarking Language Agents for Diverse, Realistic, and Long-Horizon Task Execution (ICLR 2026 · similarity 0.85)
• How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.85)
• OpenCUA: Open Foundations for Computer-Use Agents (NeurIPS 2025 · similarity 0.85)
• ShortcutsBench: A Large-Scale Real-world Benchmark for API-based Agents (ICLR 2025 · similarity 0.85)
• GitTaskBench: A Benchmark for Code Agents Solving Real-World Tasks Through Code Repository Leveraging (AAAI 2026 · similarity 0.84)
• WebPilot: A Versatile and Autonomous Multi-Agent System for Web Task Execution with Strategic Exploration (AAAI 2025 · similarity 0.84)