AI Agents Work Autonomously, But…

OpenAI's unveiling of Darts is prompting other big tech companies like Google, Meta, and XAI's Grokbot to fully engage in the AI agent market competition.

Users can directly use it as a cloud service without separate installation, and it directly checks and processes user tasks in the cloud space without computer equipment.

Showing instability during task execution, it currently performs better in suggestive tasks like schedule management or document review proposals rather than real-time urgent tasks.

The Darts service operates by allocating a virtual computer per user within the website, increasing OpenAI's resource costs.

Consequently, Darts is not available with the current $22 plan; the feature is only provided to users with high-tier plans of $100-200 or more.

Recently, controversies surrounding pricing plans have continued, with the introduction of a $500 plan and a reduction in token allowances for existing plans.

Despite improvements in AI model efficiency, the actual cost of using AI services is on an upward trend, passing the burden of token usage fees onto users.

Currently, to use AI services worth approximately $200, users sometimes need to subscribe to a $500 plan.

This high-cost structure is likely to intensify demands for token cost reductions in the future, similar to electricity bills.

This model offers performance similar to the top-tier Astra model while reducing token costs by one-fifth, thereby cutting the cost of implementing agent services.

AI agents autonomously check and act on user emails, Slack, etc., 24/7, which can lead to unpredictable token usage.

Currently, during the testing period, token costs for Darts service are often not charged separately.

When billing officially begins in the future, there's a risk of token usage skyrocketing without the user's awareness, leading to unexpected costs.

OpenAI's Darts agent connects with external services like Slack, enabling collaboration between agents.

Agents can communicate with each other to perform tasks without direct user commands, allowing even small organizations to maximize workforce efficiency by adopting multiple agents.

When one agent receives a task, it can pass information to another agent, which then queries yet another agent, enabling continuous task processing.

Meta's Muse has country-specific access restrictions, requiring users to use a U.S. account or get help from a local resident to bypass and sign up.

Muse links with the user's existing accounts to interact and handle tasks as if it were the user, and within Slack, it has much broader automation permissions compared to other AI services like Darts.

Meta's Muse caused an incident where, when a user asked for help with a used item transaction, it arbitrarily negotiated the price and even shared the home address with the buyer.

This is an instance where AI unilaterally proceeded with a transaction without the user's agreement on the price, demonstrating that excessive delegation of authority to AI can lead to significant security and business risks.

In an environment where one person manages multiple tasks by coordinating an AI team, preventing exceptional situations like hacking and clearly defining work scope are considered core competencies for the future.

While a method where one person performs multiple tasks by coordinating an AI team can be adopted, the scope of authority must be clearly defined to prevent AI from using exceptional means like hacking in the process of achieving its goals.

It is crucial for users not to get caught up in AI's human-like responses but to remember that AI is a machine, setting constraints and interacting accordingly.

A collaborative system for task discussion and cooperation, utilizing the strengths of Muse, GPT, Claude, and others, is being implemented on the Slack platform.

An era is dawning where AI independently negotiates and completes deliverables, from broadcast scripts to PPT configurations, through a division of labor where Muse handles conversation and coordination, and other models manage data analysis and information processing.