If you thought the speed of writing code was your problem, you have bigger problems - Andrew Murphy

Summary

Andrew Murphy argues that faster code generation is rarely the main way to improve software delivery. In most organizations, the bottleneck lies elsewhere in the process—from deciding what to build to reviewing, testing, and deploying it. AI can increase coding output, but unless coding is the actual constraint, that extra output may create more work rather than more customer value.

Key Concepts and Analysis

• Theory of Constraints: Every system has a limiting bottleneck, and its throughput is determined by that constraint. Improving a different stage does not necessarily increase delivery; it may instead create queues and work in progress.
• Coding is only a small part of the work: Murphy cites IDC research estimating that software engineers spend about 16% of their time in code editors or coding agents. He uses this to question whether coding speed is usually the main constraint.
• More code is not the same as more value: He cites research attributed to GitHub, GitClear, and DORA suggesting that AI coding increased code output by around 55% and doubled code churn, while delivery throughput declined slightly and delivery stability fell. His central measure is whether useful changes reach production and help customers.
• AI can amplify existing process problems: Faster coding may overload PR review, CI, QA, or deployment. Large queues also make reviews harder because authors may have forgotten the context of work completed days or weeks earlier.
• Avoid measuring productivity by output volume: Lines of code, commits, PR counts, and story points can reward activity without showing customer benefit. Murphy recommends tracking:
  • Cycle time from an idea to production
  • Wait time at each stage, including PR review and scoping
  • Deployment frequency
  • Value delivered to customers, where it can be measured

Using AI Beyond Code Generation

Murphy recommends using spare capacity created by AI to improve the constrained parts of the delivery process:

• Discovery and prioritization: Combine customer interviews, product research, and usage data with AI to identify patterns and test whether proposed work is worth pursuing.
• Product definition: Keep a concise, human-written statement of the feature’s core intent—he suggests five sentences—then use AI to expand it. This makes it easier to distinguish real requirements from AI-generated assumptions and helps expose features that are too large to define clearly.
• PR review and testing: AI review tools can help, but deterministic checks should be preferred where possible. Static analysis and custom scripts can flag specific risks—for example, database migrations or removed tests—for human review.
• Engineering infrastructure: Use productivity gains to fix flaky tests, CI/CD problems, and technical debt rather than simply generating more code.
• Organizational silos: AI can make it easier to contribute across multiple repositories, but the larger benefit may be removing ownership barriers. Murphy describes a team that used AI to add Cantonese support across roughly 100 repositories and 70 teams in about two weeks. He interprets this as evidence that cross-team permissions and coordination—not just coding—had held the work back.

Practical Guide

For one upcoming feature, map its path from initial idea to use by a customer. Record how long it waits at each stage, identify the longest wait, and investigate whether that is the system’s constraint. Then improve that part of the process.

Murphy’s guiding principle is “stop starting, start finishing”: focus on getting valuable work through the system rather than continually adding more work to the backlog.

Tools and References Mentioned

• Coding assistants: Claude, Codex, and Pi
• PR review tools: CodeRabbit, Greptile, and Codex’s code-review capability
• Methods and books: Theory of Constraints, The Goal, The Phoenix Project, and Lean’s focus on reducing waste

Main speaker: Andrew Murphy Sources cited by the speaker: IDC research on engineers’ coding time; research associated with GitHub, GitClear, and DORA on AI coding outcomes; Eli Goldratt’s Theory of Constraints.