AI Engineer Europe 2026

Software Engineering + AI = ?

Gergely Oroszswyx26:42

Loading slides…

Read the talk

Software Engineering + AI: Usage, Output, and the Work Between

Token counts can reward busywork even as coding agents change who can build software, what engineers own, and where companies invest in infrastructure.

From a talk by Gergely Orosz and swyx

When AI usage becomes a performance proxy

Should an engineer use more AI simply to make their token count go up? That is the opening problem in swyx’s conversation with Gergely Orosz, author of The Pragmatic Engineer. Token maxing can begin as enthusiastic experimentation, but a visible usage metric gives it another purpose: demonstrating that you are using the tools enough to keep your job.

Orosz describes reports from employees and managers at large technology companies. At Salesforce, he says, employees can search a tool for colleagues’ dollar spending on AI tokens. At Meta, managers have described usage as one input to performance evaluation, alongside impact, code changes, and helpful code reviews. Layoff anxiety makes that input feel consequential, even though Orosz says the Block layoffs he mentions happened independently of token spending. A low-impact engineer with low usage can be portrayed as not trying; a high-impact engineer with high usage can be portrayed as innovating. His conversations do not represent all of Meta, but they expose how an ambiguous metric can reinforce an existing judgment.

The resulting behavior is concrete. Rather than reading documentation, an engineer asks an agent to summarize it, then asks follow-up questions even when the answers are poor. Orosz reports that some engineers do this to avoid the bottom 25% or 50% of a token-count ranking. At Microsoft, he has heard complaints about autonomous agents building junk merely to increase usage. He says a Meta leaderboard was removed after an article made it look embarrassing, yet the behavior continued: removing the display did not remove the fear that someone was still measuring.

High-paying jobs raise the stakes of ignoring even a dubious signal. Orosz tentatively recalls a Salesforce minimum AI-spend target of about $175 per month, encouraging employees to reach it early in the month. What began as playful experimentation becomes compliance. The precedent is familiar: developer-productivity tools such as Velocity and Pluralsight Flow made lines of code and pull-request counts visible, and engineers at companies that rewarded those numbers learned to optimize them. Usage is an input to engineering work; making it the target can separate it from useful output.

0:280:41
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:28 · section reference included

Why leaders pushed adoption

This is a recognizable Goodhart’s-law problem: once a measure becomes a target, people adapt their behavior to satisfy it. But swyx presses the harder question. Could the push still be worthwhile overall, despite the waste it creates? To explain the incentive behind it, Orosz returns to an Amsterdam gathering of CTOs roughly six months earlier. A leader at a Dutch e-commerce company complained that engineers had AI-tool subscriptions but remained skeptical and rarely used them.

That conversation took place before Opus 4.5. Engineers working in existing codebases had tried tools such as Cursor and found them only mildly useful: the tools did not reliably perform the refactoring or find the bugs they needed help with. A representative from the Dutch National Bank described a different situation. Its engineers had a reason to experiment because understanding the technology was part of regulating it. The difference was not merely access to a subscription; it was a reason to learn.

Meanwhile, leaders saw Anthropic writing more code with Claude Code while its revenue grew. Orosz warns that this invites confusion between correlation and causation, but the managerial response is understandable: push people to use AI rather than risk standing still. In his account of Coinbase, Brian Armstrong demanded adoption within a week and subsequently fired an engineer. Orosz describes Coinbase base salaries in the $300,000–$400,000 annual range, before equity, to explain why the adoption mandate carried weight. The message reached employees through the consequences of noncompliance, not through a demonstrated improvement to their own workflow.

4:485:01
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:48 · section reference included

Compliance versus useful output

Orosz offers a provocative analogy with algorithmic hiring interviews. Big tech has used LeetCode-style screening for roughly two decades despite persistent criticism that it poorly resembles the job. In his interpretation, it selects not only for intelligence but also for willingness to satisfy an irrelevant requirement. Spending two or three months preparing for interviews before AI assistance was commonplace is one example. Someone willing to comply to obtain a desirable job may also comply with token targets to retain it.

Some engineers will turn that pressure into useful projects. But Orosz contrasts the incentives with startups, where the immediate questions tend to be what got built and what it cost. Asked whether AI is making engineering faster, he sees individual gains but remains uncertain about gains across teams. Anthropic is an example of a company moving faster; many others have not visibly followed. Accelerating an individual task does not automatically accelerate the organization that contains it. Retrofitting AI into established working practices is a separate problem.

8:018:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:01 · section reference included

Choose the right unit of productivity

The METR study of early-2025 developer productivity makes the distinction between feeling faster and finishing faster unusually clear. swyx recalls the result as roughly 20% greater perceived productivity but roughly 20% worse measured productivity. More precisely, the original report found 19% longer completion times, while participants retrospectively estimated that AI had reduced their completion times by 20%. These are changes in task duration, not symmetric changes in productivity. The experiment randomized permission to use AI on real tasks in mature repositories familiar to the developers, using early-2025 tools principally comprising Cursor Pro and Claude 3.5/3.7 Sonnet.

The sample was small: the published study involved 16 developers and 246 tasks, rather than the 30 participants Orosz recalls onstage. Orosz remembers a particularly productive outlier, whom swyx identifies as previous podcast guest Quentin Anthony. That identification is their recollection; the useful methodological point is that an average can conceal substantial differences between developers and workflows.

swyx then changes the unit of analysis. He has been enabling coding agents for nontechnical colleagues. Even if the engineer’s own coding speed barely improves, a colleague who previously had to wait for that engineer can now make progress directly. His joke about “serverless developers” points to a real organizational mechanism: removing a dependency can create value that an individual pull-request study does not measure.

Unit of analysisQuestion
Individual taskDoes a developer finish the same task sooner?
Organizational workflowCan a colleague proceed without waiting for an engineer?

Both questions matter, but answering one does not answer the other.

9:269:38
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:26 · section reference included

Learning through changing workflows

Practical fluency takes time even for experienced practitioners. Orosz recalls a 2024 conversation with Simon Willison. After a joking introduction focused on Hacker News, swyx points instead to Willison’s work on Django, blogging, and prompt injection. The relevant observation is Willison’s description of still changing his workflows after two years of experimentation. There was no manual that captured how to use the tools well.

That is uncomfortable for engineers accustomed to learning a system from its internals. Understanding compilers and assembly can directly improve low-level programming. Understanding attention and probability can explain something about a language model, but it does not by itself establish when to delegate, what context to provide, or how to recognize a useful result. Orosz’s distinction is between architectural knowledge and tool-use intuition; the latter requires repeated practice.

A successful personal workflow can also break when introduced into a team, forcing another round of learning. Orosz sees more value in teams with low ego and a willingness to revise assumptions. The invitation is to abandon unhelpful priors, not engineering experience: keep the judgment earned from building systems while testing whether an old constraint still applies.

10:4210:55
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:42 · section reference included

Broader responsibilities, smaller teams

The software engineer’s expanding role predates coding agents. Venture-funded startups have long faced pressure to grow quickly with small teams. By the mid-2010s, many expected engineers to own the code they deployed, while more traditional organizations retained dedicated DevOps teams. Orosz sees testing and operational responsibility increasingly folded into software engineering, followed by product responsibility. Companies were already hiring product engineers in 2022, before the AI shift under discussion.

AI accelerates that expansion. Even early-career engineers are expected to do more senior-like work: plan the approach, understand the business, and take responsibility beyond implementation. Smaller teams make that breadth more consequential. Orosz recounts a conversation with an engineering VP at John Deere, the roughly two-century-old tractor manufacturer, who described two-pizza teams becoming one-pizza teams partly because of the tools. The anecdote places the change beyond venture-backed software companies.

12:4112:48
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:41 · section reference included

Orchestrating agents is not people management

Does directing coding agents make everyone an engineering manager? Orosz rejects the analogy. During an informal show of hands about management experience and aspirations, swyx estimates participation at roughly 15%–20%; Orosz jokes about talking prospective managers out of it. People may pursue management for career development, compensation, or greater impact. The actual work also brings distance from the product and responsibility for interpersonal conflict, personal problems, and team dynamics. Agents do not bring that category of responsibility.

The closer comparison is a tech lead or experienced engineer directing and reviewing work. In Orosz’s conversation with DHH, the Ruby on Rails creator uses a “mech suit” metaphor: agents extend what the engineer can do while leaving the engineer in control. His image of doing seven things at once describes that feeling of amplified capability, rather than a measured throughput result.

The feedback loop also differs. Orosz illustrates management with a project involving ten reports, where the consequences of a decision may take six months to become visible. Agent work returns feedback much sooner.

ResponsibilityPeople managementAgent orchestration
Human needs and conflictCentralNot part of the agent relationship
Relationship to the productOften more indirectCan remain hands-on
Feedback on decisionsCan take monthsMuch faster

Orchestration still requires judgment, but it need not resemble running a department. Nor is maximum concurrency the goal. Orosz cites Mitchell Hashimoto as someone comfortable with a small setup; his account mentions two agents and then focuses on a single background agent. The practical point is that a manageable amount of concurrent work can be enough.

14:4214:51
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:42 · section reference included

The infrastructure hidden behind product output

Large AI budgets do not necessarily produce an immediate stream of new customer-facing features. Looking at Uber from outside, Orosz does not see a corresponding explosion of product launches. Looking inside, he sees substantial infrastructure work. Buying Cursor or Claude Code is only one part of it.

His account identifies several concrete integration points:

  • Background coding agents: custom agents connected to the company’s monorepo.
  • MCP gateway: access integrated with internal service discovery, so agents can work with the company’s services.
  • On-call tooling: operational workflows being retooled for AI assistance.
  • Code review: changes categorized by risk within the internal review system.

He names Airbnb, Intercom, Meta, Microsoft, and midsized companies as pursuing similar internal work. These investments change how engineering happens before they necessarily change what customers see.

Orosz gives three reasons for the buildout. First, internal tools provide a comparatively low-risk way to learn through use, without shipping unwanted AI features to customers. Second, enormous codebases do not fit into a context window. Custom retrieval, including basic retrieval-augmented generation, can select relevant material rather than asking a generic tool to understand everything. In that setting, he expects integration with internal context to outperform an off-the-shelf setup. Third, AI changes funding incentives: a developer-platform request that struggles to win headcount can become attractive when described as agent experience. swyx jokes that the resulting interface is just a CLI.

The consequence is substantial duplicated effort. Companies are building their own versions of similar systems, and Orosz expects the work to continue into the following year. His joke that a large company ought already to be building an MCP gateway captures how common the pattern has become. swyx connects this to the next day’s gateway and architecture talks: teams are learning from one another’s implementations because there is not yet an established textbook for the practice.

17:2517:34
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

17:25 · section reference included

Shopify’s deliberate cost of being early

It is easy to dismiss incumbent spending as waste until considering what would happen if a company such as Google executed well. Effective investment by a large incumbent can be a serious competitive threat. Shopify helps Orosz explain why accepting churn may be rational.

In his account, Farhan Thawar heard that GitHub was developing Copilot internally in 2021 and contacted Thomas Dohmke for access. The important exchange was not a purchasing negotiation: when told it was not for sale, Thawar offered a company-wide rollout in return for candid feedback. Orosz says Shopify offered feedback from 3,000 users and received Copilot about a year before wider availability. He describes Shopify as the first company to get access, with initially weak tooling and considerable churn. Subsequent early tool adoption brought generous budgets and more time spent working through bugs.

The tradeoff was deliberate: spend money and absorb instability to learn sooner. Orosz characterizes Shopify’s advantage as a few months, potentially six months, ahead of competitors. That head start can matter to a technology business; a company whose advantage lies mainly in a physical product may prefer to wait for the tools to mature. Thawar’s growing concern about costs coexists with another constraint: denying access to the best tools can make hiring the best engineers harder.

Innovation and recruiting together explain why a policy can look extravagant while still serving a business purpose. swyx adds that preparing for an interview with Shopify CTO Mikhail Parakhin exposed enough customer-facing machine learning and infrastructure to make him want to become a customer. The internal investment is most persuasive when it eventually supports something valuable outside engineering.

20:3920:53
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

20:39 · section reference included

An unexpected publishing business

The conversation closes by moving from engineering organizations to Orosz’s own business. During COVID, Uber’s layoffs hit his team, and the remaining people were redistributed when its mission no longer made sense. Low morale prompted him to take a break and write books. swyx was writing The Coding Career, and the two met through that work.

Orosz originally expected to write a book and then start a platform-engineering company based on patterns he had seen inside Uber. He points to Temporal and Chronosphere as examples of businesses emerging from that environment. swyx says that, had he not started AI Engineer, he might have started a Platform Engineer conference. Both saw value in making internal engineering knowledge available beyond the companies that developed it.

Instead, Orosz launched The Pragmatic Engineer as an experiment a year after leaving Uber. Substack was growing, and he saw room for in-depth software-engineering coverage. His first free article explained Uber’s platform/product split, a topic he says had not previously been documented publicly. A confident Twitter announcement produced demand even before the first publication. Orosz reports that 100 people prepaid $100 for an annual subscription before he published anything; within six weeks, he had 1,000 paying subscribers. He says that early subscription business replaced his former Uber base salary in Amsterdam.

22:5123:01
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

22:51 · section reference included

Protecting the work without burning out

Once he felt product-market fit, Orosz concentrated on what was working. He declined interview requests, collaborations, and podcasts to protect the article itself. The publishing commitment grew from one article to two, and for two years writing remained his focus. Only then did he step back and ask how to turn something readers wanted—and he enjoyed making—into a sustainable business.

Orosz reports working 50–60 hours even during vacations in those first two years. The work occupied his attention continuously, so sustainability required changing how it was produced. He hired Elin Bratt as the first tech-industry researcher; Jessica joined later and was present at the event. He also started a podcast roughly a year and a half before this conversation, giving an audience access to discussions he was already having.

Orosz says The Pragmatic Engineer became the number-one paid technology newsletter about four months after launch and held that position for three years. He begins to bring SemiAnalysis into the ranking discussion before the conversation turns to recognizing his role as a European technology voice. The business story ends with a concrete change in operating model: preserve the deeply researched writing that created demand, share the research workload, and expand into conversations without making permanent overwork the condition of success.

24:5124:58
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

24:51 · section reference included

Resources

From the talk

Updates since the talk

Transcript download
Full transcript

Loading transcript…