The agent era is rebuilding the stack, from the operating system to the power plant
Developers now market to agents instead of humans, a new operating system is being built for them, and venture firms argue the physical build-out should be manufactured, not constructed.
- Websites for developer tools are currently viewed more by AI agents than by human users, and companies should publish an llms.txt file for agents and serve Markdown versions of key pages when requests are identified as coming from an AI agent. m:…2402
- David Heinemeier Hansson told the a16z podcast that traditional operating systems like macOS and Windows are unfit for the age of AI agents, that Linux is uniquely suited for them because it relies on terminal interfaces, config files, and detailed error messages, and that the window for a third operating system to gain significant market share is open for the first time in 25 years. m:…1127
- Union Square Ventures argues that construction is a project while manufacturing is a product, and points to Rune, which mass-produces data centers and places them where there is excess power, and Radiant, which builds 1-megawatt microreactors that can be stacked for more wattage. m:…4585
- USV argues that AI lowers the upfront cost of manufacturing, making it viable for projects that previously could not justify factory capital expenditure, and that a product moves down the cost curve with every unit produced. m:…4585
The audience is agents
Developer marketing has a new primary reader, and it is not a person. Websites for developer tools are currently viewed more by AI agents than by human users, according to a Craft Ventures session with SpaceXSI, Braintrust and Supabase. The practical advice is direct: companies should publish an llms.txt file to provide agents with a structured sitemap of their most important documentation, and websites should serve Markdown versions of key pages when requests are identified as coming from an AI agent. m:…2402
The shift goes beyond documentation. SpaceXSI uses persistent AI agents that run on individual cloud computers to complete marketing and growth functions, and its growth and marketing team size is small relative to the company's scale due to the use of an army of bots. AI agents can be used to perform product dogfooding by signing up with their own email and reporting changes after releases. Delegating work to AI agents should focus on handing off outcomes rather than individual tasks. m:…2402
The operating rhythm changes too. Marketing in AI infrastructure is increasingly dominated by rapid response to market events rather than planned campaigns, and marketing teams should be empowered to drop planned work to respond to market events within hours. Developer tool onboarding is shifting from dashboard-based tutorials to prompt-based guidance, which is what an agent audience needs and a human audience tolerates. m:…2402
An operating system for agents
David Heinemeier Hansson, creator of Ruby on Rails, told the a16z podcast that the current duopoly of macOS and Windows is unfit for the age of AI agents. He argues that Linux is uniquely suited for AI agents because it relies on terminal interfaces, config files, and detailed error messages, and he is leveraging that to build Omachi. The window for a third operating system to gain significant market share is open for the first time in 25 years, in his view. m:…1127
The argument underneath is about cost and control. AI agents are reducing the cost and difficulty of software development to near zero, in Hansson's telling, which is what lets small teams replicate complex applications. Open-weight models are essential to prevent a duopoly in the AI industry, he argues, and Apple is at risk of becoming the AOL of the agent era due to its walled garden approach. True superintelligence may be achieved by next year, he predicts — a claim to treat as his forecast, not a fact. m:…1127
The two pieces fit together. If agents are the audience for documentation and the workers inside companies, then the interface they use most — the terminal, the config file, the error message — becomes the product surface. An operating system built for that surface is a bet that the agent, not the human, is the user that matters. m:…2402 m:…1127
Manufacturing the physical layer
Union Square Ventures makes the same product-versus-project argument about the physical world. Construction is a project, while manufacturing is a product, and AI lowers the upfront cost of manufacturing, making it viable for projects that previously could not justify factory capital expenditure. A product moves down the cost curve with every unit produced, enabling venture-scale growth, which is the property investors actually care about. m:…4585
The examples are all in the AI build-out. Rune mass-produces data centers and places them where there is excess power. Radiant makes small nuclear power plants in a factory and ships them, building 1-megawatt microreactors that can be stacked for more wattage. Isembard automates CNC factories for aerospace, space, and defense industries. m:…4585
USV extends the logic to the bottlenecks. It identifies drug manufacturing as a potential area for transition from construction to engineering-based manufacturing, and grid hardware like transformers and substations as a bottleneck suitable for product-based manufacturing. It suggests standardized robotics kits could replace custom site integration for tasks in critical industries. The pattern is the same in each case: take a one-off build, turn it into a repeatable product, and let the cost curve do the work. m:…4585
What it means for investors
The three threads add up to one thesis: the agent era rewards whoever builds for agents as the primary user, at every layer. Developer tools that are not readable by agents lose the audience that now reads them. An operating system that is not built for terminal-driven agents leaves the door open for a third entrant for the first time in 25 years. And physical infrastructure that is still built as one-off construction projects cannot ride a cost curve. m:…2402 m:…1127
The risk is the usual one: a thesis about a cost curve is not a guarantee that the curve arrives on schedule. But the direction is consistent across all three sources this week, and it is the direction venture money follows — toward products that get cheaper with every unit, whether the unit is a data center, a reactor, or an operating system. m:…4585 m:…1127
What is still unresolved
- Whether a third operating system can actually win share against the macOS and Windows duopoly
- Whether factory-built data centers and microreactors reach production volumes on venture timelines
- Whether agent-readable documentation becomes a standard or stays a developer-tool niche
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.