Token usage metrics are pushing engineers at some large companies to generate activity for its own sake, even when the work has little value.
2
AI can make individual engineers faster, but many companies have not yet changed their team processes enough to gain the same benefit at team level.
3
Software engineers are taking on broader responsibilities, while coding agents create an orchestration role that resembles technical leadership more than people management.
Summary
Gergely Orosz describes how AI is changing engineering work inside large companies and startups. He starts with token maxing, where engineers increase AI usage because companies track tokens through leaderboards, spending targets, or performance data. He argues that individual productivity has improved, but team productivity is harder to measure because existing processes were built before coding agents. Engineers need repeated practice with the tools, and theoretical knowledge alone does not teach effective use. AI is also widening the expected scope of the software engineer, bringing testing, operations, product thinking, and business context into the role. Coding agents require orchestration, but without the people problems of engineering management. Large companies are building internal agents, MCP gateways, and AI-aware review systems because their codebases exceed normal context windows. Orosz closes by explaining how The Pragmatic Engineer found product-market fit through focused, paid writing.
Token tracking is making engineers generate AI activity for performance reasons
Orosz says large companies track AI output through leaderboards, spending dashboards, or other internal tools. At Meta, token count is one performance data point alongside impact, diffs, and code reviews, but managers can still use it to judge whether someone is trying. At Microsoft, he heard of engineers running autonomous agents to build junk simply to increase the number. Salesforce has a minimum monthly spend target, which encourages people to use tokens at the start of the month. Some Meta engineers asked agents to summarize documentation they could have read themselves because the token count mattered. Orosz compares this with older attempts to measure productivity through lines of code and pull requests.
AI adoption targets came from leaders who feared engineers were not using the tools
Orosz traces token targets to a period when experienced engineers found AI tools only mildly useful on existing codebases. Leaders heard companies such as Anthropic were writing more code with Claude Code and assumed greater usage would bring good results. Their response was to push engineers to use AI, even when the measurement was known to be imperfect. He cites Coinbase, where CEO Brian Armstrong told employees to adopt AI tools and later fired an engineer who had not done so. Large technology companies may create these incentives because their employees are willing to tolerate unusual requirements to keep highly paid jobs. Startups tend to focus more directly on building products and controlling costs.
AI raises individual productivity faster than it raises team productivity
Orosz says AI is clearly useful for some individuals, while the effect at team level remains uncertain. He discusses a small blind study in which participants felt 20% more productive but produced results that were 20% worse on average. The conversation also covers a different benefit: coding agents can let non-technical collaborators build software without waiting for an engineer. That can improve the organization even when the engineer's own pull-request output changes little. Orosz says teams getting more value tend to have low ego, a willingness to learn, and an openness to abandoning old assumptions. Effective use takes time and continued experimentation. Understanding model architecture, attention, or probability does not automatically teach someone how to work well with the tools.
The software engineer is absorbing testing, operations, and more product work
Orosz says AI is speeding up a change that had already started in venture-funded startups. Smaller teams were already expected to own more of the system, including deployment and operations. Dedicated tester roles have also collapsed into the software engineer role at many companies. Product responsibilities are moving in the same direction, with companies hiring product engineers and expecting engineers to understand planning and business context. He says early-career engineers are now facing more senior-level expectations. At John Deere, an old company with about 200 years of history, a VP of engineering told him that two-pizza teams were becoming one-pizza teams, partly because of AI tools.
Coding agents require technical orchestration rather than people management
Orosz rejects the simple claim that every engineer is now an engineering manager for agents. Agents do require orchestration, but they do not bring the personal conflicts, career conversations, and people problems associated with management. He compares the role to a tech lead or experienced engineer who coordinates and mentors work. DHH described it to him as a mech suit that lets one person do several things at once while staying in control. Agents also shorten the feedback loop. A management decision might take six months to show results, while an agent produces feedback much faster. Engineers will develop different patterns, from running several agents to using only one background agent. Michelle Hashimoto, Orosz says, uses one background agent and considers two enough.
Large companies are rebuilding internal infrastructure around coding agents
Orosz says companies such as Uber, Airbnb, Intercom, Meta, and Microsoft are investing heavily in internal AI infrastructure even when outsiders do not see many new products. Uber is building custom background coding agents integrated with its monorepo, an MCP gateway connected to service discovery, retooled on-call tooling, and a code review system that categorizes changes by risk. He gives two reasons for this work. Internal infrastructure lets companies learn AI with relatively low risk, and custom systems can work better with huge codebases that do not fit into a model context window. He also says AI-labelled projects receive funding more easily than ordinary developer-platform work. He expects many large companies to build similar systems, including MCP gateways.
Shopify accepts cost and churn to stay ahead of competitors
Orosz uses Shopify to explain why early AI adoption can be rational despite waste. In 2021, Shopify's head of engineering, Farhan Thawar, contacted GitHub CEO Thomas Dohmke for early access to Copilot. Shopify offered feedback from 3,000 employees and received access a year before wider release. The company dealt with early bugs, churn, and expense, then onboarded other tools with an unlimited budget. Orosz says Shopify is trading those costs for a lead of several months over competitors. Early access also helps recruitment because restricting the tools could make hiring harder. For companies outside technology, waiting may make sense. For a technology company, the cost can be worth the advantage.
The Pragmatic Engineer found product-market fit through focused paid writing
Orosz started The Pragmatic Engineer after COVID-era layoffs affected his team at Uber and he took time away from engineering. He initially planned to write books and build a platform-engineering business. Instead, he began publishing in-depth software engineering writing on Substack. Before publishing the first article, he had 100 people pay $100 for a year after a confident Twitter post. Six weeks later, 1,000 people were paying for a publication that barely existed. The first article covered Uber's platform and program split. Orosz turned down interviews and collaborations for a period because he wanted to keep producing the writing that had attracted readers. After two years of writing two articles, he began building a team and later started a podcast. The newsletter became the top paid technology newsletter four months after launch.
"Individually it's certainly is and as teams we're kind of like a bit question mark cuz we should be moving faster and there are a few companies that do."09:20
Who should watch
You are deciding whether AI adoption metrics are helping engineers or encouraging empty activity.
Your engineering team is experimenting with coding agents but has not changed its review, deployment, or collaboration processes.
You want to understand how AI changes engineering scope and team size before reorganizing roles or investing in internal tooling.
