YouTube Review · July 2026

Slow Down to Speed Up AI Software Engineering

Slow down to speed up: AI and software engineering is Gergely Orosz's Craft Conference keynote for The Pragmatic Engineer. The video is useful because it refuses the clean version of the AI-coding story. The claim is not simply that agents write more code, nor that AI coding is fake. The sharper claim is that generated output has moved faster than many teams' ability to verify, review, integrate, and trust the work.

The talk opens with the Meta Instagram account-recovery incident as a warning about AI systems with authority over sensitive workflows. The Verge separately reported that Meta said it had fixed an issue where an AI support chatbot could be manipulated into linking a new email address to someone else's Instagram account. Orosz then uses the incident to make an engineering-culture point: when organizations push AI use faster than security, review, and ownership can adapt, the failure is not only a model failure. It is a management failure.

The most important Spiralist signal is the verification bottleneck. The talk surveys OpenAI, Cursor, Google, Meta, Uber, startups, and traditional companies, but the recurring pattern is the same: code generation becomes cheap, while product judgment, review capacity, tests, architecture, observability, incident response, and senior engineering attention remain scarce. A team can produce more diffs and still get slower at building reliable software if the review surface grows beyond the trust surface.

Orosz's discussion of tokenmaxxing gives the review its governance hook. The Pragmatic Engineer's related Pulse article describes engineers and teams facing rising token spend, usage measurement, internal pressure to adopt AI, budget limits, and early attempts to connect spend to outcomes. The danger is that token volume becomes the new activity metric. If a company rewards usage before it measures useful shipped work, people will optimize the meter. The token counter then becomes a dashboard version of the mythical man-month: countable, comparable, and not the same as progress.

For this archive, the video belongs beside The Mythical Man-Month. Brooks warned that adding people to a late software project can add communication and integration work faster than useful output. Orosz is documenting the AI-era version: adding agents to a software process can add review, provenance, integration, and confidence work faster than useful product change. The unit changed from man-month to token, diff, agent run, or generated pull request. The coordination problem survived.

External evidence supports caution rather than a single verdict. METR's 2025 randomized study found experienced open-source developers working on familiar repositories took longer with then-current AI tools, while also warning that the result was a time-specific snapshot and not a universal claim. DORA's 2024 report found AI adoption associated with individual productivity, flow, and satisfaction, but also with tradeoffs around delivery stability and throughput. Those findings can coexist. AI coding is not one thing; task type, codebase familiarity, review culture, test quality, domain risk, model version, tool scaffolding, and organizational incentives all change the result.

The practical governance lesson is narrow and useful: do not let AI-generated work outrun the evidence layer. A serious AI-software workflow should preserve the task, prompt or instruction, model/tool context, changed files, source material, tests run, tests missing, reviewer, blast radius, rollback path, and human owner. It should also measure outcomes that matter: defect rate, cycle time after review, incident load, maintainability, user value, security findings, and whether human engineers still understand the system.

Uncertainty should stay visible. This is a conference keynote and industry-analysis video, not a neutral benchmark or audited incident report. Some claims are based on Orosz's sources, experience, and private conversations; some are supported by public reporting; some remain insider interpretation. Treat the video as strong evidence of the software-industry mood in June 2026: AI coding is real enough to change workflows, expensive enough to force budget governance, and risky enough that the old engineering disciplines matter more, not less.

Sources

Return to YouTube