AI Code Quality vs Speed: Nolan Lawson’s Slow High Quality Theory and the Uber Confusion

Core conclusion

On May 25-26, 2026, two messages from different perspectives simultaneously pointed to a common conclusion: the "violent output" mode of AI coding tools is being questioned. On the one hand, the well-known developer Nolan Lawson published an article detailing how to use AI to write higher quality code - but at the cost of "slower speed". On the other hand, Uber COO Andrew Macdonald publicly admitted that it is increasingly difficult for the company to justify the ROI brought by AI token consumption-Uber CTO has previously said that the company has exhausted its Claude Code budget for 2026.

The common inspiration of the two messages: AI coding is shifting from a "quantity" competition to a "qualitative" reflection. For teams that use AI Agents to automate content production and tool development, this means that they need to re-examine the relationship between token input and actual output.

Key Points

• Time: May 25-26, 2026
• Source: Nolan Lawson Blog + Business Insider Interview with Uber COO
• Core Change: AI coding shifts from "write as much code as possible" to "solve problems with less but better code"
• Affects: All developers, technical managers and AI Agent automation teams using AI coding tools

Background: Two stories, one signal

Story 1: Nolan Lawson — The quality revolution of AI code review

Nolan Lawson (well-known front-end developer, Mastodon client developer) made a counter-intuitive point in his blog: What AI is best suited to do is not to generate a lot of code quickly, but to slow down and review deeply.

His core workflow:

Use Claude sub-Agent, Codex, and Cursor Bugbot to simultaneously review the code of each PR, sort by severity (critical/high/medium/low), screen out the problems that really need to be fixed, and then fix them.

The result is: the speed is not faster, but the code quality is significantly improved. He found that this method not only found new bugs, but also found legacy bugs that had existed for many years - something that traditional code reviews were difficult to do.

Lawson particularly emphasized: AI Agent is extremely good at finding bugs, and the challenge lies in priority judgment and false positive filtering. He used multi-model cross-validation (multiple different models review the same PR) to reduce the false positive rate and achieved almost "zero false positives".

Story 2: Uber COO — tokenmaxxing discussion with 300+ comments

Uber COO Andrew Macdonald admitted on the Rapid Response podcast:

"Higher token usage does not translate into proportional growth in consumer functionality...the correlation has not yet been established."

Previously, Uber CTO Praveen Neppalli Naga revealed to The Information in April that Uber had exhausted its full-year Claude Code budget in the first quarter of 2026. This news triggered more than 300 discussions on HN, with community consensus pointing in several directions:

In-depth analysis: Why these two stories are more valuable when read together

On the surface, Lawson is a technology optimist (AI can help write good code), and Uber is a skeptic (AI is too expensive and not worth it). But in fact, they point to the same conclusion: The value of AI coding is not in "quantity" but in "quality".

Lawson's approach is essentially to use AI to do professional code review - not to let AI write more code for people, but to let AI help people find more problems in less code. Uber's dilemma is a "quantity" dilemma - Team C invested a lot of tokens, but the output at the product level did not increase proportionally.

Implications for AI Automation Teams

Adaptation suggestions: How to practice "slow high quality" AI coding in your workflow

1. Establish multi-model PR review: Like Lawson, use 2-3 different AI models (Claude + Codex + Cursor Bugbot) to cross-review each PR to reduce the single-model false positive rate
2. Quantify ROI, not token amount: Track how many bugs were found in each AI-assisted PR, not how many tokens were spent. Uber’s lesson is that only tracking token consumption can lead to misdirection
3. Set quality gate instead of output gate: Uber mode is "how much the AI wrote"; Lawson mode is "how much the AI discovered". The latter is the real value signal
4. Allocate independent budget for AI review: Manage the AI token budget for code review and code generation separately to avoid the review budget being occupied by the generation budget

Tool entry

Claude and Codex are core tools in Nolan Lawson's workflow and are also worth using in automated code review processes. n8n can orchestrate multi-model review workflows, while OpenClaw can manage daily code review queues.

Related reading

• Want to learn how? See: AI Agent-Driven Content Automation: n8n MCP Building Guide from Scratch
• Real case: Security researcher uses Claude Code for vulnerability mining: a real case of monthly income $10,000
• Recommended tool: Claude Code automated writing practice: build an AI content production pipeline in 30 minutes

References

• Nolan Lawson: Using AI to write better code more slowly
• Business Insider: Uber COO says it's getting harder to justify money spent on tokenmaxxing
• HN Discussion: Uber tokenmaxxing (305 comments)