From Atomic Actions to Standard Operating Procedures: Iterative Tool Optimization for Self-Evolving LLM Agents
06:00 · July 9, 2026 · arXiv cs.AI RSS
Tool utilization enables Large Language Model (LLM) agents to interact with the real world and resolve complex tasks. However, existing agent frameworks predominantly rely on static toolsets composed of granular atomic actions (e.g., basic file I/O or single-turn search), which forces agents to reinvent low-level logic for every recurring workflow, leading to increased reasoning overhead and failure rates. In this study, we propose that agents can achieve self-evolution by synthesizing these atomic actions into reusable Standard Operating Procedures (SOPs), which function as callable higher-order tools that encapsulate multi-step logic. We further introduce EvoSOP, a framework that empowers agents to extract SOPs from execution trajectories and iteratively optimize the toolset through a systematic lifecycle of construction, merging, evaluation, and pruning. Extensive experiments demonstrate that EvoSOP significantly boosts task success rates while substantially reducing the number of interaction rounds compared to baselines. Our analysis also reveals that iterative tool optimization fosters reliable and efficient tool-use patterns, providing a scalable pathway for the development of self-evolving agents.
Summary
Existing LLM agent frameworks typically equip models with static collections of low-level atomic actions, such as individual file operations or single-turn searches. Because these primitives contain no higher-level structure, agents must reconstruct the same multi-step sequences each time a recurring workflow appears, which inflates reasoning cost and raises the probability of failure.
The paper proposes that agents can overcome this limitation by distilling successful execution traces into reusable Standard Operating Procedures. Each SOP functions as a callable, higher-order tool that encapsulates a verified sequence of atomic steps, allowing the agent to invoke complex behavior with a single reference rather than re-deriving the underlying logic.
EvoSOP implements this idea through a closed-loop optimization process. From raw trajectories the system first constructs candidate SOPs, then merges semantically similar procedures, evaluates their reliability on held-out tasks, and finally prunes those that add little value or introduce errors. The resulting toolset is fed back to the agent for subsequent episodes, producing a progressively refined library of procedures.
Empirical results show that the iterative refinement raises task-completion rates while lowering the average number of environment interactions required. The authors attribute the gains to the emergence of stable, task-appropriate tool-use patterns that reduce both redundant reasoning and exposure to intermediate failure points.
Why it matters
This research is highly relevant for Dutch AI researchers and enterprise developers building autonomous agents, as it offers a novel method to reduce reasoning overhead and API costs while improving reliability. The transition from static tools to self-evolving SOPs aligns well with the Dutch market's focus on scalable, efficient AI automation for SMEs.
