AI agents and the software meltdown – Crypto News

They are both part of the same story, just about the only story in tech: artificial intelligence, and the rise of autonomous agents.

Semiconductor stocks that are connected to AI data centers are the center of the bullish action, with no signs of the investment train slowing down. In the past few weeks, Amazon.com projected 2026 capital expenditures of around $200 billion, and Alphabet’s Google estimated up to $185 billion. Meta Platforms said it would spend as much as $135 billion, and it doesn’t even have a cloud unit that rents out AI servers like the other two companies. Adding in Microsoft, Oracle, and CoreWeave, we could see three-quarters of a trillion dollars invested in AI data centers.

Since OpenAI’s ChatGPT lit this prairie fire in November 2022, the semiconductor ETF is up around 190%.

The flip side is the negative narrative on software and information services—that many of these companies will not survive the AI age. The first victims, like the educational service Chegg, are already struggling and trying to pivot to something new. Chegg stock is down 98% since ChatGPT’s debut. Many are asking, “Who’s next?”

Investors, for now, are throwing the babies out with the bathwater. Big software names such as Salesforce are trading at their lowest price/earnings ratios ever; Salesforce shares are down by 29% this year. Investors used to have confidence in software companies’ long-term cash-flow generation. They no longer do.

Questions about enterprise software first popped up as early as 2022. Not only can large language models, or LLMs, like ChatGPT, Anthropic’s Claude, and Alphabet’s Gemini speak human languages very well, they are even better with computer programming languages. This means more software can be produced at a new very low cost, devaluing current enterprise software makers, or even replacing them altogether.

To date, coding is the most successful commercial use of these LLMs. The LLM makers saw the traction they were getting with software developers and focused a lot of attention on coding skills, advancing them with each generation.

A major milestone was the debut of Claude Code from Anthropic about a year ago. This took AI coding capabilities to a new level with agents—software that can use LLMs to accomplish a complex series of tasks from a simple prompt, automating major chunks of the development process. OpenAI, Google, and others followed, and now some developers see themselves as conductors of an orchestra of coding agents that are simultaneously working on different parts of an application at the same time. Despite reported negative side effects like increased cognitive load and exhaustion, few developers deny that coding agents work much faster than humans can.

By November, there was a large cohort of technically-minded people who had figured out how to use coding agents, and they were prime candidates for something bigger. Originally released as a weekend project named ClawdBot, OpenClaw is a desktop agent with powerful abilities to work with websites, data, software, and communications, with added integrations on Apple Mac computers. Because of that and other benefits, the humble Mac mini has become the weapon of choice for techies to run their coding and desktop agents, using the LLMs as the foundation.

OpenClaw is free, open-source software that is difficult for the mass majority of people to use. I spent close to two days getting it to work on my Mac mini, but eventually I was able to prompt it to do a multistep task that takes me about 20 minutes to complete a few times a month.

The combination of great powers combined with hacked-together agent software and LLMs that can still get things very wrong is a security nightmare. Cybersecurity companies like CrowdStrike Holdings are giving customers tools to detect when OpenClaw is running on corporate devices and disable it. For security’s sake, I ran OpenClaw without any access to my data or passwords, but most aren’t being that careful, because depriving agents of these privileges makes them far less useful.

“Trying to make OpenClaw fully safe to use is a lost cause,” said security start-up Aikido in a February blog post. “It’s only useful when it’s dangerous.”

Please do not install OpenClaw on your own computer unless you fully understand all the risks.

Anthropic released its own desktop agent this year. Called Claude Cowork, it includes prewritten prompts for different use cases like basic legal work, which amplified the noise around agents, causing the iShares software ETF to decline 5% that day.

And that brings us back to the software apocalypse. Not only are agents out there writing software much faster than humans can, but now we see a demonstration of how powerful they can be in automating all sorts of knowledge work on a computer. It has even led to a rethinking of what “software” means in a world in which we primarily interface with computing devices through prompts, not mouse clicks.

OpenClaw—open-source software hacked together with the aid of AI and accessible only to techies—has been a huge contributor to the collapse of software and information services stocks. This week, OpenClaw’s developer, Peter Steinberger, accepted a job at OpenAI. That’s the AI Age in a nutshell.

As I wrote in an earlier edition of Tech Trader, I believe that the software meltdown is overblown, and that there are buying opportunities on the horizon. But for now, that doesn’t count. The narrative is the only thing that matters.

Write to Adam Levine at adam.levine@barrons.com