This paper discusses strategies for reducing token costs associated with AI agents, particularly within the OpenClaw open-source platform, through the introduction of a new feature called "Lobster."

OpenClaw is presented as an open-source platform designed to automate AI agents, capable of connecting with messaging services like Telegram and Discord to perform scheduled tasks such as email checking, news aggregation, and data analysis. The core problem addressed is the significant and often unforeseen token consumption when these AI agents execute repetitive tasks. Tokens are defined as the unit of usage, akin to the number of characters an AI processes (reads and writes). The paper explains that AI agents, when performing tasks like "collecting and organizing popular projects from GitHub," engage in complex internal processes: determining data sources, reading responses, deciding on data formatting, and handling retries. Each of these "thinking" steps involves an AI model call, consuming tokens. This leads to inefficiency because the AI redundantly "thinks" through the same steps for identical, repetitive tasks, analogous to constantly using a map application for a daily commute.

To address this inefficiency, OpenClaw launched Lobster on February 17, 2026. Lobster is designed to offload repetitive, non-creative tasks from the AI. Its core methodology involves executing tasks based on a pre-defined "configuration file," which functions like a recipe. This file explicitly dictates the sequence of operations (e.g., "Step 1: Fetch data from GitHub, Step 2: Sort by popularity, Step 3: Save results"). By leveraging this configuration, Lobster executes these steps sequentially and mechanically, without requiring AI inference at each stage. This allows the AI agent to receive pre-processed data, enabling it to focus solely on creative tasks such as analysis, interpretation, and content generation.

The paper highlights a hybrid approach where Lobster handles data collection, parsing, and storage—tasks that are highly pattern-based and repetitive—while the AI is reserved for complex analysis, interpretation, and strategic decision-making based on the data provided by Lobster. For example, in a "Tech Trend Collection" task that involves scraping five different websites, the previous method required 2-3 AI calls per site, totaling over 10 AI calls for a single collection, as the AI determined the next steps. With Lobster, the data collection is performed with a single, pre-defined sequence, reducing AI involvement to just 1-2 calls for analysis of the collected results.

Empirical results from applying Lobster to 9 scheduled tasks demonstrated a significant reduction in token usage. Over a two-week comparative period, the total token consumption (measured by the volume of characters processed by the AI) decreased by 30-40%. Beyond cost savings, Lobster also improved operational efficiency by substantially reducing task timeouts. Previously, AI's "thinking time" in intermediate steps could extend the overall task duration, leading to timeouts. Lobster's direct, sequential execution of pre-defined commands is significantly faster, mitigating this issue. Additional benefits include enhanced consistency in output structure, as Lobster always adheres to the defined sequence, and improved resilience through features like resuming tasks from where they stopped.

The paper concludes that Lobster is most beneficial for users actively employing AI agents for automation, particularly those with five or more daily scheduled tasks, repetitive data collection patterns, or monthly token expenditures exceeding $10 (approximately 14,000 KRW). It emphasizes that the core principle for cost reduction is to assign tasks that do not require AI's "thinking" capabilities to dedicated tools like Lobster, thereby optimizing the allocation of AI resources for higher-value, creative processes. While initial time investment is required for setting up configuration files, this investment yields continuous savings over time.