Agents make many software projects economically viable because they reduce the cost of encoding business knowledge and process logic.
2
The easiest useful agents research information, expose knowledge that already exists, automate repetitive support work, and leave final decisions with people.
3
AI engineers can build an application layer above increasingly commoditized models, while making software usable by agents through APIs, CLIs, sandboxes, and safer infrastructure.
Summary
Malte Ubl argues that AI engineering will follow web development as a mainstream software discipline. Agents make software viable in areas where traditional applications were too expensive to build, so the amount of software companies create should grow even as development gets faster. He describes practical agent patterns, including 24/7 support work, research before a human decision, keeping company information current, and removing repetitive tasks from jobs. Vercel uses agents to route sales inquiries, research abuse reports, and answer support questions. Ubl also says software itself is increasingly used by agents. More than 60% of Vercel.com page views came from AI agents during the previous seven days, and APIs and CLIs are becoming more important than dashboard workflows. He expects model companies to commoditize, leaving application-layer engineers with more control over where business value and new AI techniques are created.
Agents make previously unaffordable software worth building
Ubl describes agents as a new kind of software because they can automate work that traditional applications were too expensive to encode. Conventional systems require engineers to turn business knowledge into many rules and if statements. That cost meant companies left much software unwritten, even when it should have existed. Agents reduce the effort enough to make more of those projects viable. Ubl connects this to the choice between buying SaaS and building internally. He says companies are increasingly choosing to make software themselves, and frames the change as an economic test of how much software demand expands when software becomes cheaper to produce.
Simple agent patterns can create value without changing the whole process
Ubl says teams do not need to start with ambitious autonomous systems. One useful pattern is automating a job that currently runs during working hours so it can operate continuously. Another is compressed research: an agent handles the research between a business event and a human decision. The process stays the same, but research that took 30 minutes might take five. If it happens 100,000 times per year, the savings are substantial without increasing the risk or changing who makes the decision. He presents this as practical work that companies can ship today.
Vercel uses agents to research sales and abuse reports
Vercel's contact-sales workflow sends messages to an agent. About 75% of the time, the agent identifies a support request and routes it to support. Otherwise, it researches the person's LinkedIn profile, looks up the company, estimates its size, and routes the inquiry to the right person. A human still reviews the result. Vercel uses the same pattern for abuse reports. An agent investigates whether a website is abusive and gathers the relevant information, while a professional makes the final decision. Ubl says this removes research work that previously took a person around 15 minutes.
Agents can make existing company information usable
Many companies already have the information needed for a task, but it is scattered across issue trackers, Slack messages, and meeting recordings. Ubl suggests building agents that gather this material and keep operational records current. When a manager asks for updates, the agent could already have assembled them instead of starting a new manual research task. This pattern does not require a new source of data. It makes information that already exists practical to use. The same approach can apply wherever a business process depends on facts that are present in company systems but are difficult for people to collect.
Removing repetitive support work can improve the work left to people
Ubl describes Vercel's in-house support agent, which has a 90% deflection rate. In most cases, it helps the customer in real time rather than sending the issue through another support path. The result, he says, was a large increase in job satisfaction for the support team. People no longer spent much of their time on routine questions such as rejected credit cards. They could focus on unusual cases and customers who needed deeper help. Ubl presents the removal of boring work as a worthwhile goal when deciding where to build agents.
Software must become usable by agents as well as people
The application layer changes when agents become software users. Ubl says that more than 60% of page views on Vercel.com in the previous seven days came from AI agents. Platform use is also moving from people clicking through dashboards toward APIs and CLIs. He now asks for the CLI or automation path whenever someone proposes a feature with a user interface. He expects infrastructure practices to change too, because developers may have less attachment to code they did not write themselves. Applications are also becoming agents, which creates demand for infrastructure such as sandboxes.
Agent security needs separation between the harness and generated code
Ubl describes the current security situation as dangerous because agents can run code and interact with systems in ways that are hard to control. He argues that many popular agent harnesses have the wrong architecture because they combine the environment running the harness with the environment where generated code executes. Separating those environments is important. He notes that Anthropic's newer agent product had adopted this separation. His broader point is that the field is still early, and future paradigm shifts will require engineers to reconsider assumptions about how agents, sandboxes, and infrastructure should fit together.
The application layer can grow even when model providers change
Ubl says AI engineers can build a stable application layer even as different models take turns being better. He expects model companies to commoditize, with competition pushing infrastructure and model prices lower. In that outcome, agents built by application engineers create the business value, and new AI engineering techniques emerge above the models. He contrasts this with a future where one major lab controls the best models and application developers become forward-deployed engineers for that provider. Ubl considers the commoditized-model outcome more likely and says it gives AI engineers more power over the products and processes they build.
"In many ways I think AI engineering is the legitimate successor to web development as a really mainstream discipline of engineering that will shape the next decade of software development."02:29
Who should watch
You are deciding which internal processes are worth automating and want agent ideas that leave human decisions in place.
Your product has a dashboard-first workflow, but agents are starting to use your service through APIs or CLIs.
You are building agent infrastructure and need to think about sandboxes, execution boundaries, and changing model providers.
