One year after Dario Amodei said AI agents would arrive at scale in 2025 and AI would write virtually all code within twelve months, the results are real inside Anthropic, partial across the broader industry, and nearly absent in the high-stakes professional environments where they matter most.
The prediction was not vague. Speaking in early 2025, Anthropic CEO Dario Amodei forecast that AI agents would see large-scale deployment within the year and that 90% of software code would be AI-written by the end of 2025. At Davos in January 2026, he told the World Economic Forum that AI could replace most of what software engineers do within six to twelve months. Those are not hedged, carefully qualified statements. They are deadlines, and today is the day the internet is holding him to them.
Inside Anthropic, the prediction has held. The company confirmed in early 2026 that over 90% of code for new Claude models and features is now written autonomously by AI agents, with human engineers acting as architects and security auditors rather than primary developers. Claude Code's internal success rate on the company's most challenging tasks doubled in the final four months of 2025, while the number of human interventions per session fell from 5.4 to 3.3, according to Anthropic's own published research on agent autonomy. Boris Cherny, the engineer who built Claude Code, publicly predicted in February 2026 that the job title of software engineer would begin fading this year. These are not external claims. They are disclosures from the people building the systems.
Zoom out beyond Anthropic's walls and the picture changes significantly. Fortune reported at Davos that few C-suite leaders outside the AI sector concurred with Amodei's job displacement timeline. At most software companies, AI-written code sits between 25% and 40%, roughly half what Amodei projected industry-wide. In legal, finance, and office administration roles, Anthropic's own capability data reveals a gap that the company itself has published: Claude can theoretically handle 94% of tasks in the computer and mathematics category but is currently covering just 33% in real-world deployments. Legal and finance roles show similar gulfs between what the model can do in a benchmark and what enterprises are actually delegating to it.
The reasons are well-documented and not technical. Hallucination rates on extended agentic tasks remain high enough that errors compound over long autonomous sessions. Legal liability has not been resolved: no framework yet exists for attributing professional responsibility when an autonomous agent makes a consequential mistake. And enterprise procurement cycles, which typically take twelve to eighteen months even for straightforward software, have not moved at the speed the AI labs anticipated. Anthropic's January 2026 launch of Claude Cowork, the company's most ambitious agentic product to date, was explicitly designed to bridge this gap by giving the model access to local file systems, browser control, and multi-step workflow orchestration. But as Axios noted at launch, the claims are the company's own, and independent enterprise adoption data for autonomous delegation at scale remains thin.
The Harder Question Behind the Benchmark
The trending debate on Reddit and X today is framed as a question of whether Anthropic missed its deadline. That framing is too simple. What actually happened is more structurally important: the technical capability arrived faster than the institutional infrastructure required to use it. The models can do the work. The workflows, liability frameworks, professional standards, and organizational habits required to hand that work to machines have not caught up. This is not a new dynamic in technology. The internet was capable of transforming retail by 2000. Amazon did not overtake Walmart until 2021. The lag is not evidence that the prediction was wrong. It is evidence that technological capability and economic adoption travel at different speeds, and the AI industry, accustomed to measuring progress in benchmark months, has consistently underestimated how long the second part takes.
Dario Amodei made his predictions about AI employees in a context where Anthropic's internal experience was the reference point, and inside Anthropic the predictions came true. The error was assuming that the rest of the world would adopt at the same pace as an AI-native company with no legacy workflows, no external liability exposure, and direct access to the models before anyone else. The industry is not behind where Amodei thought it would be because the technology failed. It is behind because organizations are harder to change than models are to improve, and that gap will narrow on its own timeline, not the CEO's.
Also read: Elon Musk's universal high income proposal is bigger than UBI and more complicated than he admits • DeepSeek V4 arrives one year after its market shock and raises the stakes in the US-China AI race • AI will almost certainly create the world's first trillionaire and the race is already underway
