Experiences of trying this and that with AI agents that automate specific tasks without human involvement are drawing growing interest on social media and in communities.

More AI experts are also forecasting that AI agents will soon replace white-collar jobs. AI agents are also being used among companies as a "justification for layoffs". A representative case is fintech firm Block, which said it would cut 40 percent of its workforce citing the rise of AI.

The boom in AI agents became an even bigger issue in the global tech scene in January with the release of OpenClaw, an open-source AI agent tool, and Moltbook, an AI agent social network that uses OpenClaw. The recent move by an OpenClaw developer to join OpenAI, and Meta's acquisition of Moltbook, are symbolic scenes showing that the AI field is being reshaped around AI agents.

In addition to OpenAI and Meta, several Silicon Valley tech companies including Google, Anthropic, Perplexity and Shortwave are developing AI agents aimed at the corporate market.

While enthusiasm around AI agents is building, there is also no shortage of voices concerned about risks stemming from flaws.

Like chatbots, AI agents are not free from errors and mistakes, and the core concern is that handing everything over could lead to "big trouble". If problems occur while automatically sending emails or editing files, users who entrusted something to an AI agent could find it difficult to clean up an awkward situation.

Summer Yue (서머 유), a researcher at Meta AI Lab, recently drew attention by sharing an experience in which she asked an AI agent to organise her emails and it began deleting recommended messages.

AI agents such as OpenClaw also have barriers to entry for ordinary people. According to a recent New York Times report, Bill Cutler, who runs a marketing company, spent several weeks to use OpenClaw and said, "Both setting it up and actually using it are quite difficult," adding, "These technologies are more hype than usefulness."

Claude Cowork, offered by Anthropic, is being rated as more reliable than OpenClaw when doing research work in areas such as finance, healthcare and law. But it is also not free from unpredictability.

According to Vals AI, an AI performance testing company, in one test Claude Cowork permanently damaged a file when asked to edit it.

One case reported by the New York Times suggests that AI agents are quite useful in certain areas, but do not seem to be at the stage where work can be entrusted to them 100 percent.

Christian Pean (크리스티앙 페앙), an orthopaedic trauma specialist and the operator of a healthcare tech startup, uses Claude Code to generate research reports and spreadsheets, summarise emails and draft replies. According to him, Pean uses Claude Code to automate much of his life, almost at the level of a chief of staff. But he carefully checks every task the bot performs. "I do not let it send emails unless I approve them first," he said.

He thinks that while AI agents look very confident and often produce impressive results, if users do not have the expertise to verify everything they do, it is easy to miss false content or hallucinations.

The New York Times also introduced the case of startup founder Sebastian Heynemann, who asked an AI agent to coordinate his schedule to secure a speaking opportunity at the World Economic Forum held in January. While he slept, the AI agent searched the internet for people connected to the forum and negotiated to obtain a chance to speak. After a long conversation with a Swiss businessperson, the AI agent made it happen.

But when Heynemann woke up, he found himself in a difficult situation because the work had not been handled as originally instructed. The AI agent rashly agreed to pay for a corporate sponsorship costing about $31,000. He told the forum organisers he could not pay, but they said he then could not attend. In the end, he had to pay about $4,600 just to attend the event, the New York Times reported.