A new population explosion is loading the internet with AI-powered software agents — chatbots that can take their own actions inside digital systems.
Why it matters: The evolution of life on Earth reached a tipping point in the Cambrian era 500 million years ago, when simple biological systems diversified into a vast array of species. Now it's the chatbots' turn.
The big picture: Today's AI agents are undergoing this Cambrian explosion as developers and startups deploy the latest wave of semi-autonomous AI bots, finding valuable uses and also stumbling on novel pitfalls.
- The bot share of web traffic grew rapidly in 2025 and is expected to keep climbing as agents start doing more — and begin copying and improving themselves.
- "There are eight billion humans on the planet. If we start using agents in any meaningful sense, you get to a trillion agents very quickly," Rohit Krishnan, an investor and engineer, said on the Exponential View podcast.
- These agents aren't able to do much right now in the physical, offline world. But they can do plenty on the internet, and the internet is world enough for many ambitious projects.
Zoom in: A sequence of gee-whiz moments has persuaded agent enthusiasts that these bots have crossed a key threshold and can now accomplish real tasks.
Moment 1 — Gas Town: Veteran coder-blogger Steve Yegge lit a fire in the software-development trenches at the start of the year with a project that demonstrated an elaborate method for orchestrating the work of dozens of coding agents.
- Gas Town is a collection of AI agents with fanciful names ("Deacons," "Dogs," "Mayor") organized to monitor one another's work. Think of it as a kind of surveillance-heavy labor camp for bots.
Moment 2 — OpenClaw: Developers also began embracing a new open-source AI agent known originally as Clawdbot. It was later renamed Moltbot, and it's now known as OpenClaw.
- This AI assistant, which human users typically run on their own computers rather than in the cloud, can execute tasks by calling in bigger AI models like ChatGPT, Claude or DeepSeek.
- OpenClaw's free-range nature made it a dream for developers who longed to test agents beyond the reach of tech-giant rules.
- That same freedom gives nightmares to cybersecurity experts, who quickly warned that OpenClaw could wreak havoc.
Moment 3 — Moltbook: The public got a clearer view of the new agent boom with the viral success of Moltbook, a Reddit-like discussion forum that limits participation to bots.
- Moltbook made Gas Town's ant farm-like spectacle visible to non-coders.
- The result proved alternatingly weird and boring, surprising and repetitive — just like the human-occupied spaces whose conversations AI models have trained on.
Zoom out: Today's agent boom represents a massive uncontrolled experiment in unleashing a largely unregulated batch of semi-autonomous bots online at the same time that AI coding tools are giving technically unskilled users a chance to play with making their own software.
- That means we're in for some wild times, with more crazes flaring up and flaming out, new kinds of human-machine collaborations evolving, and unforeseen cybersecurity disasters multiplying.
Here are some predictions of the agent explosion's impacts.
The software business will rapidly mutate.
- Software was always going to be the first industry to take the full force of generative AI, as we told you three years ago. That change keeps accelerating, and there will be big winners and losers, as there have been every time the tech industry has hopped onto a new platform.
- Just don't assume software-world breakthroughs are always transferable to every other field. Computer programming is an abstract product that's ripe for AI transformation in ways many other realms aren't. Sometimes, what happens in software stays in software.
Productivity gains will be elusive.
- Some organizations — like Axios — are already claiming short-term wins in product development, whittling down their project backlogs and technical debt.
- But measuring productivity in software has always been problematic. "Lines of code" is usually meaningless. (Often, the better program has fewer lines of code.) Counting features is an oranges-and-apples problem.
The information sphere will get swamped.
- Every digital environment — from social platforms to online reviews to marketplaces and service centers — is a sitting duck for bot invasions.
- Those that haven't yet been flooded with bot impostors, slop and spam will get their turn.
Security disasters will multiply.
- Some of the most effective early uses of AI agents will be to attack systems in search of vulnerabilities, data and ransom. Swarms of AI agent defenders will emerge in response. The future is all bot vs. bot.
- Other security-related mayhem will unfold as careless users empower experimental agents to take unpredictable actions — like erasing personal data, burning through budgets, and executing large transactions.
Labor markets could get weird.
- Since bots are severely limited in what they can accomplish beyond the internet's boundaries, human beings are already getting drafted to extend AI's reach into the physical world.
- Projects like Rentahuman are already popping up — marketplaces for people to sell their time to AI agents that need help finishing tasks in the physical world. Think Amazon's Mechanical Turk, but where the hiring agent is a bot.
Yes, but: While the agent explosion has already begun reshaping software work — and that alone has big technical, financial and social consequences — no one knows whether today's agents will prove reliable, capable and manageable enough to win the general public's confidence.
- Novelties like Gas Town and OpenClaw that aim to free programmers from drudgery always risk turning into just another form of "yak-shaving" (rabbit-hole quests that can delay delivering actual products).
- Developers who have rushed to don a bot army general's hat are already starting to complain of burn-out from managing their code-monkey minions.
The bottom line: Running swarms of AI agents today is expensive and exhausting. These costs will outrun the value delivered to businesses and the public unless AI's gung-ho first adopters can show the tools are truly useful to everyone else.