
OpenAI is about to unleash a Swarm of multi-agent systems on enterprises that just might cost decision-makers their jobs.
The company’s experimental framework designed to orchestrate networks of AI agents signals the next step in a shift toward building AI agents that are capable of handling tasks without human interaction. But industry leaders and AI ethicists fear the technology will lead to job displacement, security risks and potential bias.
OpenAI insists Swarm is not an official product.
“Swarm is not an official OpenAI product. Think of it more like a cookbook. It’s experimental code for building simple agents. It’s not meant for production and won’t be maintained by us,” OpenAI researcher Shyamal Anadkat wrote in a post on X.
Business operations could be revolutionized by Swarm, leading to enhanced efficiency and freeing employees to focus on deeper-thinking projects. A network of specialized AI agents, for example, might work together to collect and analyze market data, adjust strategies based on marketing trends, identify sales leads and provide customer support — with little to no human involvement.
Indeed, an open-source project, called “OpenAI Agent Swarm Project: Hierarchical Autonomous Agent Swarms (HOS),” which is gauging Swarm’s potential, includes a hierarchy of AI agents with distinct roles and responsibilities. How it will be governed remains another matter, however.
News of Swarm has rekindled debates about the ethical use of advanced AI systems. Without safeguards, autonomous agents might lapse into biased and unfair conclusions that adversely impact individuals and society, according to security experts.
Companies remain dubious in adopting agentic AI. More than half (55%) of survey respondents have trust-related issues, ranging from data privacy and ethical issues to industry regulations and legal standards, according to a new report from Forum VC.
At the same time, the emergence of Swarm comes weeks after OpenAI announced o1, its new “reasoning model” that brings STEM-like skills to chatbots. Imagine a swarm of o1 agents at a workplace.
To further muddle matters, scientists at Apple Inc. including AI expert Samy Bengio, question the logical capabilities of today’s large language models.
Researchers tested proprietary OpenAI models as well as open source models such as Llama, Gemma and Mistral. What they found, as published on arXiv, was that even leading models like OpenAI GPT-4o and o1 don’t use logic, but merely mimic patterns.
Though the OpenAI o1 series received top scores on many benchmarks, it falls short on performance fluctuations and makes “silly mistakes,” the researchers concluded. The finding was supported by another recent study.
“Overall, we found no evidence of formal reasoning in the language models,” lead researcher Mehrdad Farajtabar said in the report. “Their behavior is better explained by sophisticated pattern matching.”
Scaling data, parameters and compute would lead to better pattern matchers, but “not necessarily better reasoners,” he added.