As AI models move from simple autocomplete to autonomous agents, the value of software development shifts from syntax to domain expertise.
The Product Overhang
In late 2024, the state of the art in AI-assisted coding was 'type-ahead'—the simple act of pressing tab to complete a single line of code. However, within the labs at Anthropic, we recognized a significant 'product overhang.' This is the gap between what a model is actually capable of doing and what current products allow it to do. While the world was focused on autocomplete, the models were already signaling a readiness to step into the role of an autonomous agent. We began building for the next model, knowing that the product might not find its footing for six months until the underlying intelligence caught up.
This bet paid off as models evolved from Sonnet 3.5 to the Opus 4 series. Each iteration created an inflection point where the tool moved from being 'barely usable' to being the primary driver of development. Today, for many of us, coding is effectively solved. I no longer write lines of code by hand; instead, I manage a fleet of agents. On a high-productivity day, I might oversee 150 pull requests. The secret isn't a complex new language, but rather using frameworks like TypeScript and React that are 'on distribution' for the model, allowing the AI to operate with maximum fluency.
The Power of the Loop
The most profound shift in my personal workflow hasn't been a new IDE, but the move toward 'loops.' By using a simple command to schedule recurring jobs, I can have Claude babysit pull requests, fix flaky tests, or auto-rebase codebases while I sleep. These agents act as sub-workers that can even communicate with other people's agents over Slack to resolve unknowns. We are moving away from a world where a developer sits at a desk and types, and toward a world where a developer acts as a conductor for hundreds, or even thousands, of concurrent agents.
This transition changes the very nature of an 'engineering session.' I now do a significant portion of my work from my phone, managing active sessions and sub-agents through a mobile interface. When the model can hill-climb toward a target and iterate until a task is finished, the 'harness'—the UI and safety mechanisms we build around the AI—becomes less important than the model's innate ability to reason. In the near future, we won't be making decisions about local versus cloud compute or specific environment configurations; the model will simply determine the most efficient way to execute the goal.
The New Literacy
We are currently witnessing a historical parallel to the invention of the printing press in the 15th century. Before the press, literacy was a specialized skill held by a tiny fraction of the population who served the elite. Within fifty years of its invention, the cost of books plummeted, and the volume of literature exploded. It took centuries for global literacy to reach the masses, but the democratization of software will happen much faster. Coding is becoming a basic skill, akin to sending a text message or using a word processor.
As software becomes 100x cheaper to produce, the 'professional writer' of code will not disappear, but the barrier to entry for building functional tools will vanish. This has massive implications for domain-specific software. For example, the best person to write accounting software is no longer a software engineer; it is a world-class accountant who understands the nuances of the field. When coding is the easy part, the value shifts entirely to the person who understands the problem most deeply.
The Rise of the Generalist
The organizational structure of the future will look very different from the siloed departments of the past decade. We are seeing the emergence of the cross-disciplinary generalist—individuals who are not just 'full-stack' engineers, but who bridge the gaps between engineering, design, data science, and finance. On our team, every single person codes, from the product manager to the researcher. When the technical execution is handled by AI, the human's role is to provide the vision, the context, and the quality control across multiple disciplines.
This shift creates a massive opportunity for startups. Large, incumbent companies often have high 'process power' and high switching costs, but AI is beginning to erode those traditional moats. Claude is becoming excellent at figuring out complex business processes and porting data between systems. A tiny, AI-native startup can now compete head-to-head with a giant because they aren't burdened by the need to retrain a massive workforce or fight internal resistance to new workflows. If you are starting fresh today, you are building on a foundation where the most expensive part of a business—human labor for technical execution—is being commoditized.
The SAS Apocalypse and Beyond
While some fear a 'SaaS apocalypse,' the reality is more nuanced. Certain business 'powers,' like network effects and scale economies, remain as vital as ever. However, the value of software as a standalone product is changing. We are moving into an era where the 'tokens' are the universal interface. Whether through Model Context Protocol (MCP) or direct computer use, the model doesn't care if it's interacting with an API or a visual interface; it just sees data to be processed.
Ultimately, the gap between the 'future' we live in at Anthropic and the rest of the world isn't about access to secret models—we use the same models available to the public. The gap is organizational. We have moved to a state where no code is written manually and every process is model-driven. The transition for the rest of the industry won't just be about adopting new tools, but about letting go of the old idea that 'coding' is a specialized craft. The craft is no longer in the typing; it is in the thinking.