How AI is changing Software Engineering: A Conversation with Gergely Orosz, @pragmaticengineer

April 21, 202626:4216,848 views

Summary

In this fireside chat at AI Engineer Europe 2026, Gergely Orosz — former Uber and Skyscanner engineer and author of The Pragmatic Engineer newsletter (the #1 software/AI engineering newsletter on Substack) — discusses how AI is transforming software engineering culture, productivity, and team structure, drawing on extensive first-hand reporting from engineers at Meta, Microsoft, Salesforce, Shopify, and other large tech companies.

Orosz opens with a deep dive into "token maxing" — a phenomenon where engineers at large companies artificially inflate their AI token usage to meet internal metrics or avoid performance flags. At Meta, token consumption appears as one data point in performance evaluations, leading engineers to ask agents to summarize documentation they could read themselves, solely to boost token counts. At Microsoft, engineers run autonomous agents to generate junk output for the same reason. Salesforce has set a minimum target of $175/month in AI spend per employee. Orosz traces this behavior to a broader leadership push: CTOs, frustrated that experienced engineers were skeptical of AI tools on existing codebases pre-Opus 4.5, began measuring adoption to force usage. He draws a parallel to Leetcode-style interviews — a practice also criticized as irrelevant to the actual job — noting that Big Tech repeatedly selects for people willing to "put up with absolute nonsense to keep the job."

On the productivity question, Orosz is candid. He references the Metr study (approximately 30 participants) in which engineers felt 20% more productive but demonstrated 20% less productivity on average, while one outlier was genuinely highly productive. His own theory is that the real productivity unlock is enabling non-technical collaborators to write code themselves — removing the dependency on engineering queues — rather than making individual engineers dramatically faster.

Regarding the evolving engineering role, Orosz argues that the software engineer has been absorbing adjacent roles (QA, DevOps) for years, and AI is accelerating the absorption of product thinking as well. At a 200-year-old company like John Deere, two-pizza teams are becoming one-pizza teams. He pushes back on the framing that engineers are now "engineering managers for AI," arguing that orchestrating agents is more like being a tech lead or senior engineer — you have the orchestration responsibility without the people management, personal conflict, and slow feedback loops that make management genuinely hard. He cites DHH's description of it feeling like a "mech suit" — doing seven things at once while staying in control.

On large tech internal infrastructure, Orosz details how companies like Uber are rebuilding internal tooling from scratch: custom background coding agents integrated into mono-repos, MCP gateways integrated into service discovery, AI-assisted on-call tooling, and risk-categorized code review systems. He explains the logic: codebases too large to fit in a context window are better served by custom solutions than off-the-shelf vendors, and developer platform teams can secure headcount more easily by rebranding work as "agent experience."

Shopify's AI adoption story is highlighted as a strategic case study. In 2021, Head of Engineering Farhan Thawar secured early access to GitHub Copilot — before it was commercially available — by offering 3,000 engineers' feedback in exchange, accepting significant churn and expense to be six months ahead of competitors. Orosz frames this as rational: for companies where technology is core, the churn cost of being at the frontier is worth the recruitment and competitive advantage.

Orosz closes with his own journey building The Pragmatic Engineer, which reached the #1 paid technology newsletter on Substack within four months of launch and stayed there for three years, growing from 100 paying subscribers in the first week to a business requiring a team, now including a podcast launched 18 months prior.

Key Takeaways

•Token count metrics at Meta, Microsoft, and Salesforce are driving perverse behaviors — engineers are using AI unproductively just to hit numbers; teams should measure outcome quality, not token volume.
•The Metr study (~30 participants) found engineers felt 20% more productive with AI but demonstrated 20% less productivity — expectations and reality diverge significantly; instrument actual output metrics.
•The real AI productivity unlock may be enabling non-coding collaborators to write code themselves, removing bottlenecks, rather than making individual engineers dramatically faster.
•Getting good at AI tooling takes a long time and continuous iteration; understanding transformer architecture or attention mechanisms does not translate to practical prompt/agent skill.
•Companies with very large codebases that cannot fit in a context window should build custom AI solutions and MCP integrations rather than relying solely on off-the-shelf tools.
•If your large company does not already have an MCP gateway integrated into service discovery, you are likely already behind your peers who are all building the same infrastructure.
•Orchestrating AI agents is not equivalent to engineering management — it lacks the people drama and slow feedback loops, and is closer to a senior tech lead or mech-suit model (DHH framing).
•Shopify's strategy of accepting high churn and expense to be six months ahead of competitors on AI tooling is rational for tech-core companies, where the recruitment and competitive benefits outweigh costs.
•When engineering platform teams struggle to get headcount, reframing work as 'agent experience' secures budget — a real dynamic worth understanding when lobbying for AI infrastructure investment.
Product-market fit for technical content can be validated before publishing: Gergely Orosz got 100 paid subscribers at $100/year before publishing a single article, reaching 1,000 within 6 weeks.

Transcript

Your game. All right. I I going to assume most of you Show of hands, who subscribes to Pragmatic Engineer? Oh my god. >> Wow. Uh he is he needs no introduction then. Let's get right into it. Um What is token maxing and should everyone here be doing it? So I I heard about token maxing a week ago or like week and a half ago first and you know, some people have been doing it for longer and I tweeted about it I think 3 days ago saying, "Oh, there's this token maxing." And again, you you see it on so...