The Bottleneck Moved From Code Generation to Review and Judgment

# The Bottleneck Moved From Code Generation to Review and Judgment When AI makes code generation nearly free, the scarce work doesn't disappear — it moves to the two ends generation never touched: *problem-translation* upstream and *review/verification* downstream. Typing was never the constraint. So the engineer's value shifts from output to consequence — "it's not so much about what you have built, it's more about what has changed because you built it" — and one of the biggest strengths becomes "the ability to say no" to features and code that shouldn't exist. Engineers become co-owners of the business problem with PMs, not order-takers behind a wall. The downstream bottleneck is concrete and human: senior reviewers. Staff engineers "become the bottleneck after some point because the amount of volume that they see, they cannot effectively context switch between all of these different PRs to actually give a meaningful review." The fixes follow from that: move feedback *upfront* (iterate on design/architecture before code exists, not after), and hand deterministic checks to deterministic tools — an ESLint rule "is just so much better than an LM can ever do… cheaper, faster, deterministic." Review quality becomes a tracked metric (CodeRabbit's ~64% suggestion-acceptance, up from the low 20s) rather than a vibe.  ## Why This Matters For **[[AI & the Future of Knowledge Work MOC]]**, this is the engineering-specific instance of the general amplification claim: [[3. Resources/Permanent Notes/AI Amplifies Existing Bottlenecks Rather Than Creating Capability]] says AI 10x's whatever was already the constraint — and in software the constraint is review, not typing. So "AI made us faster at coding" amplifies the wrong step and floods the real bottleneck. It also explains *why* the maker/checker discipline matters more under AI, not less: [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] is the structural answer to a review bottleneck — if the human reviewer can't context-switch across the volume, the verification has to be a separate, automated, evidence-producing step, not a tired human skim. ## Counterpoints to Hold - **"Move feedback upfront" assumes you know the design upfront.** For genuinely exploratory work, the design *is* discovered by writing code; front-loading review can ossify a wrong architecture early. - **Deterministic-tools-beat-LLMs is domain-bound.** True for lint/format/type checks; false for "does this match the product intent," which is exactly the judgment the note says is now scarce. The split between the two isn't always clean. - **Saying "no" requires standing.** "Ability to say no" is a strength only for engineers with the seniority to be heard; for juniors it can read as obstruction. The skill is unevenly available. ## Links - [[3. Resources/Permanent Notes/Template-Presence Is the Diagnostic of Automatability]] — extends: code generation follows reusable patterns, while review remains difficult because its judgment cannot be reduced to the same stable template - [[3. Resources/Permanent Notes/Implementation Is Cheap So Taste Becomes the Bottleneck]] — extends: review verifies generated work, while taste decides which cheap implementations are worth pursuing at all. - [[Auditing the Build Not the Install Is What Makes Personal-Scale Coverage Possible]] — applies-in: build deduplication makes automated scanning cheap enough that human proof and judgment become the residual constraint. - [[Abstracting Drudgery Re-Opens Creative Work to Non-Specialists]] — extends: abstraction widens access to production while review and judgment become the next scarce specialization. - [[AI Erodes the WordPress Plugin Value Props Pushing Monetization Toward SaaS and Services]] — applies-in: AI can generate plugin updates, but review and integration judgment remain the costly part that services can monetize. - [[AI Is a Building Block You Build From Not a Tool You Add to a Workflow]] — extends: rebuilding work around AI moves human effort from generation into the review and judgment layer. - [[3. Resources/Permanent Notes/AI Amplifies Existing Bottlenecks Rather Than Creating Capability]] — the general law; this is its engineering case. - [[3. Resources/Permanent Notes/Clean Code Is the Human-AI Interface]] — what makes the now-binding review step tractable: enforceable standards the reviewer (and the agent) can lean on. - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — the verification has to be a separate checker, since the human reviewer is the bottleneck. - [[3. Resources/Permanent Notes/AI Compresses the Knowledge-Worker Pipeline From Both Ends]] — what a review-and-judgment premium does to the junior-to-senior pipeline. - [[3. Resources/Permanent Notes/Code Review Is a Human Knowledge-Transfer Ritual Not Just a Quality Gate]] — the complement: review isn't just the scarce *constraint*, it's the scarce *formation ritual*, so you can't just automate the bottleneck away without losing the judgment it produces. - [[3. Resources/Permanent Notes/AI Finds Exploit Chains Faster Than Humans Can Audit Them]] — the security-research case: an agent finds a working chain in hours while human verification and explanation take longer - [[3. Resources/Literature Notes/My AI-Assisted Content System That Keeps Me in the Loop with Kyle Balmer]] — the content-production case: automated drafts make human rewrite and editorial judgment the binding step - [[3. Resources/Permanent Notes/Encoding What Good Looks Like Is What Makes Agent Verification Trustworthy]] — applies-in: a checked-in quality standard turns scarce review judgment into criteria an agent and checker can apply repeatedly. - [[3. Resources/Permanent Notes/Heavy AI Use Removes the Fatigue Signal That Tells You to Stop]] — applies-in: sustained orchestration and review replace effortful generation with work whose cognitive stopping cue is easier to miss - [[3. Resources/Permanent Notes/Route AI Models by Temperament Not Just Cost Tier]] — applies-in: a model that exposes uncertainty lowers the scarce review burden more than one that hides gaps behind fluent completion - [[3. Resources/Permanent Notes/Sampling a Fraction of a Source Produces False Authority Over the Whole]] — evidence-for: scarce human judgment must detect when a confident answer rests on insufficient source coverage - [[3. Resources/Permanent Notes/Spec-Driven Development Turns Throwaway Prompts Into Durable System Memory]] — applies-in: executable specifications preserve review judgments so they do not have to be reconstructed for every generated implementation - [[3. Resources/Readwise/Articles/It's the AI Gold Rush]] — added 2026-09-05: the WP Minute podcast's own phrasing of this note's claim — "syntax is solved" but composition and judging which generated option is right is not — arrived at independently in a WordPress-development podcast rather than the engineering-org sources this note was built from. - [[3. Resources/Readwise/Articles/AI Struggles A Developer’s Story]] — independent convergence: a maintainer describes the contribution-specific version of this claim (external AI-assisted PRs are often cheaper to rewrite than to review, because generation moved to the contributor's tool while review stayed with the maintainer), reached without reference to this note - [[3. Resources/Permanent Notes/A Career Track Is a Reversible Seat Not an Identity]] — the career consequence: if judgment is the scarce half, the IC seat is an allocation decision rather than a step down. ## Open Questions - If review is the bottleneck and juniors generate the volume, who builds the judgment that senior review depends on once the junior rungs are automated? - Can "review quality" be measured without gaming it — does an acceptance-rate metric just train reviewers to accept more? ## Source - [Reader: What are engineers getting paid for now?](https://read.readwise.io/read/01kv3sb6xfcgmvdsskztr9emam) — value is what changed, not what was built; co-owners with PMs; the ability to say no - [Reader: What to do when there's too much code to review](https://read.readwise.io/read/01kv3t68q9rekxv4k4xvpbegr9) — staff engineers as the context-switch bottleneck; feedback upfront; ESLint beats the LLM; 64% acceptance - See also: [[AI & the Future of Knowledge Work MOC]]