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

Key Concepts

• Token Maxing: The practice of artificially inflating AI token usage to meet internal company metrics, performance quotas, or to avoid being flagged as a low-performer.
• AI Infrastructure (Infra): The internal systems, custom agents, and gateways (e.g., MCP gateways) companies build to integrate AI into their specific codebases.
• Productivity Paradox: The phenomenon where engineers feel more productive using AI, but objective data may show a decline in output or quality.
• Orchestration vs. Management: The shift in the software engineer's role from traditional coding to managing and directing autonomous AI agents.
• Priors: Preconceived notions or established workflows that engineers must be willing to abandon to effectively leverage AI.

---

1. The Phenomenon of "Token Maxing"

"Token maxing" has emerged as a culturally strange byproduct of corporate AI adoption. At large companies like Meta, Microsoft, and Salesforce, AI usage is often tracked via leaderboards or spend-tracking tools.

• Drivers: Employees fear that low token usage will be interpreted as a lack of innovation or effort during performance reviews.
• Behaviors: Engineers are resorting to "junk" tasks—such as asking agents to summarize documentation they could read themselves—solely to increase their usage numbers.
• Corporate Response: Some companies have implemented minimum spend targets (e.g., $175/month at Salesforce). Others, like Meta, have removed public leaderboards after internal backlash, though the culture of "maxing" persists due to job security anxieties.

2. Strategic AI Adoption and Infrastructure

Large tech companies are investing heavily in custom AI infrastructure rather than relying solely on off-the-shelf tools.

• Custom Solutions: Companies like Uber and Airbnb are building internal "MCP (Model Context Protocol) gateways" and custom background agents integrated into their monorepos.
• Why Build vs. Buy:
  1. Context Window Limitations: Off-the-shelf models cannot handle the massive, proprietary codebases of large enterprises.
  2. Risk Mitigation: Internal tooling provides a controlled environment to experiment with AI without exposing sensitive code to public vendors.
  3. Funding Incentives: "Agent experience" projects are currently easier to secure headcount for than traditional developer platform improvements.
• The Shopify Case Study: Shopify secured early access to GitHub Copilot by trading honest, high-volume feedback for exclusive access. They accepted the "churn" and expense of early adoption to maintain a competitive edge in recruitment and developer velocity.

3. The Changing Role of the Software Engineer

The industry is shifting toward smaller, more autonomous teams where the role of the engineer is expanding.

• Role Collapse: The traditional distinctions between QA, DevOps, and Product roles are collapsing into the "Software Engineer" role.
• The "Mech Suit" Analogy: Rather than becoming "Engineering Managers," engineers are becoming "orchestrators." Using AI feels less like managing people and more like operating a "mech suit" that allows one person to perform the work of several.
• Higher Expectations: Even early-career engineers are now expected to have a deeper understanding of business impact and system architecture, as AI handles the rote coding tasks.

4. Productivity and Learning

• The Productivity Paradox: A cited study showed that while engineers felt 20% more productive with AI, they were objectively 20% less productive. However, the speaker notes that enabling non-technical collaborators to code via agents can create a massive organizational unlock, effectively creating "serverless developers."
• The Learning Curve: There is no "manual" for AI. Understanding the underlying architecture (attention mechanisms, probability) does not necessarily make one better at using the tools. Success requires a "low ego" approach—being willing to leave behind "priors" (old ways of working) and constantly iterating on workflows.

5. Notable Quotes

• "It selects for the person who's smart and willing to put up with absolute [BS] to get the job... this person will be putting up with [BS] that makes absolutely no sense to keep the job." — On why high-performing engineers tolerate arbitrary metrics like token maxing.
• "It's more like a mech suit where you have like you can do seven things at once, you can do a lot faster and you're in control." — On the nature of managing AI agents.
• "If you're at a large company and you're not already building an MCP gateway, what are you even doing?" — On the current state of internal AI infrastructure.

Synthesis

The current state of AI in the enterprise is characterized by a tension between genuine innovation and performative metrics. While "token maxing" represents a dysfunctional response to corporate measurement, the underlying trend—building custom AI infrastructure and shifting toward "orchestration" roles—is a rational, albeit expensive, attempt to stay competitive. The most successful teams are those that treat AI as a tool for organizational leverage rather than just an individual productivity booster, prioritizing open-mindedness and rapid iteration over rigid adherence to legacy engineering practices.