From Atomic Actions to Standard Operating Procedures: Iterative Tool Optimization for Self-Evolving LLM Agents

2026-07-09 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, extended

Summary

EvoSOP is a novel framework designed by researchers from Renmin University of China and Alibaba Group, enabling Large Language Model (LLM) agents to self-evolve through iterative tool optimization. It addresses the limitations of static, atomic toolsets by synthesizing multi-step logic into reusable Standard Operating Procedures (SOPs). EvoSOP employs a systematic lifecycle of construction, merging, evaluation, and pruning to refine the agent's toolset. Experiments on ACEBench and Tau2Bench demonstrate that EvoSOP significantly boosts task success rates, yielding gains from 2.5% to 13.4% over baselines, while substantially reducing interaction rounds. The framework is model-agnostic, requires no parametric LLM updates, and maintains a compact toolset, typically fewer than 10 SOPs, achieving approximately 80% task success.

Key takeaway

For AI Scientists and ML Engineers developing robust LLM agents, adopting an iterative tool optimization framework like EvoSOP is crucial. You should move beyond one-shot tool creation by implementing continuous lifecycle management for your agent's toolset. This approach, encompassing SOP construction, merging, evaluation, and pruning, will significantly enhance task success rates and reduce reasoning overhead, preventing toolset bloat and fostering more reliable, efficient agent behavior in complex, stateful environments.

Key insights

LLM agents can self-evolve by iteratively optimizing toolsets through SOP synthesis and systematic management.

Principles

• Iterative tool optimization fosters reliable, efficient tool-use patterns.
• Continuous toolset refinement prevents performance degradation from bloat.
• SOPs reduce cognitive load by encapsulating multi-step workflows.

Method

EvoSOP iteratively constructs SOPs from execution trajectories, merges redundant routines, evaluates their real-world utility, and prunes ineffective tools in a continuous optimization loop.

In practice

• Synthesize recurring action sequences into callable SOPs with executable code.
• Merge functionally overlapping SOPs to maintain a compact, generalized toolset.
• Prune SOPs exhibiting high error rates or low utility based on empirical performance.

Topics

• LLM Agents
• Tool Optimization
• Standard Operating Procedures
• Self-Evolving AI
• Iterative Learning
• Agent Frameworks

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

Related on AIssential

• Learning to Construct Practical Agentic Systems — Modular agentic systems with learned fixed workflows offer superior cost and accuracy over dynamic planning.
• EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer — EvoAgentBench evaluates LLM agent self-evolution by isolating procedural knowledge transfer, moving beyond simple task accuracy or memory retention.
• AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning — LLM agents can achieve dynamic tool selection and superior generalization by optimizing tool preferences via Plackett–Luce ranking.
• Evoflux: Inference-Time Evolution of Executable Tool Workflows for Compact Agents — Evoflux repairs failed tool workflows for compact LMs via inference-time evolutionary search, significantly boosting execution feasibility.
• Stop coding AI: Use Runtime Topological Self-Assembly (UC, DeepMind) — LLMs can optimize discrete symbolic graphs to discover new algorithms and dynamically architect multi-agent systems.
• Tools as Continuous Flow for Evolving Agentic Reasoning — FlowAgent reconceptualizes LLM tool chaining as continuous trajectory generation for robust, generalized reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.