Read More
Anthropic recently launched Claude Mythos, a powerful new frontier AI model that is significantly stronger than its flagship Claude Opus 4.6. However, it is not open to the public and only a few selected partners, such as Apple, Google, Microsoft, and Nvidia, can currently use it.
ADVERTISEMENT
SCROLL TO CONTINUE WITH CONTENT
Mythos is extremely good at finding unknown security holes (zero-day vulnerabilities) in software. It can discover bugs in operating systems, web browsers, and open-source software, including some that had remained hidden for up to 27 years.
I want to use this as an illustration to discuss the future of AI agents, a currently hot topic in AI. Today, many AI companies build “agents,” which are smart helper programs that use code, tools, and structures (collectively called scaffolding) built upon large language models to make them more capable of handling complex, real-world tasks on their own. For example, an AI coding agent can read software code, test ideas, and find bugs with very little human intervention. This makes AI agents useful for many areas like coding, cybersecurity, finance and medicine, or tasks where one prompt is not enough. Agents are clearly valuable right now because they are turning today’s LLMs into practical workers that help businesses save time and money.
However, Mythos also shows why AI agents can become obsolete in the long run. Its strong skills in finding vulnerabilities did not come from heavy extra scaffolding or agentic tools. Instead, these abilities emerged naturally during training as the model improved at reading, understanding, and modifying code. This is what Anthropic described as “emergent” behavior. When base models grow smarter, they develop “agentic” powers inside themselves and therefore no longer need lots of outside rules or wrappers, as known as agents.
As a result, pure wrapper-agent companies may lose their edge. Experts predict over 40 percent of agent projects could be dropped or reworked due to escalating costs, unclear business value, and inadequate risk controls. While agents will still matter in specialized areas like finance and medicine, the future belongs to powerful base models that can think and act like agents on their own. Agents are helpful tools today, but smart base models like Mythos may make many of them unnecessary tomorrow.
Allen Au is a tech startup founder, AI architect, and YouTuber
