The Page Was Never the Bottleneck
When Aldus PageMaker and the Apple LaserWriter arrived in 1985, producing a paginated document stopped being the hard part. Newsrooms that responded by scaling production rather than editorial review discovered an awkward asymmetry: a cheap composition tool amplifies an unverified sentence exactly as efficiently as a verified one. Output rose. Reliability did not follow. Software engineering is now running the same experiment, and the industry is misreading the result because it is still measuring the wrong unit.
What the 1968 Conference Actually Diagnosed
The useful precedent is not the compiler, awkwardly, but the institutional reaction to it. FORTRAN and COBOL had removed the need to hand-write instruction sequences by the late 1950s, yet delivery performance did not improve correspondingly; specification errors simply migrated to a higher level of abstraction. At the 1968 NATO Software Engineering Conference, the response was to stop treating programming as an individual craft and start treating it as an engineering problem requiring configuration management, decomposition and verification. The lesson for this cycle is narrow: an order-of-magnitude improvement in production capability delivers no matching gain in reliability until the management layer is rebuilt.
The Pull Request Is the Congested Intersection
Generation has been commoditized; review has not. An agent can produce a plausible multi-file change in the time a senior engineer spends reading the ticket, so the queue has moved from the editor to the reviewer. Diff-based review — the discipline's default control — assumes a human author who considered each line and a change set small enough to hold in working memory. Neither assumption survives machine-scale output. The practical response is executable specification: property-based tests, mutation testing to establish that a suite can actually fail, and contract tests at service boundaries. Teams that add agents without adding verification throughput have not increased capacity. They have increased work in progress and hidden it behind a green CI status.
The Productivity Claim Does Not Survive Its Own Control Group
The pessimistic reading deserves its own stress test. In METR's randomized controlled trial, published 10 July 2025, 16 experienced open-source developers working on 246 tasks from repositories they already knew took 19% longer when AI tools were permitted — while still believing afterward that the tools had helped them. That is a real finding and an inconvenient one. It is not a general law. The sample concentrates on the people with least to gain from assistance: authors who already hold the codebase in their heads. Peng and colleagues' separate 2023 controlled experiment found that participants using GitHub Copilot completed a bounded JavaScript task 55.8% faster than the control group. The defensible inference is not that AI slows developers. It is that gains depend on task boundedness, context familiarity and where verification cost sits — and that self-reported productivity is not a control group.
The Clock Starts at Awareness, Not at the Patch
Article 14 of Regulation (EU) 2024/2847 requires an early warning within 24 hours of a manufacturer becoming aware of an actively exploited vulnerability or severe incident, with a substantive notification within 72 hours. The consequence rarely discussed is that the clock attaches to awareness, not to remediation readiness. A team that discovers a defect on a Friday afternoon now owns a regulatory obligation before it owns a fix. That favors organizations with continuous dependency visibility and penalizes those whose bill of materials is a quarterly export.
Generated code adds a second-order problem. When ten thousand repositories are patched by the same model against the same idiom, defects stop being independent; they become a monoculture of reasoning rather than of libraries. Provenance must therefore extend beyond the commit hash to model version, retrieval set and tool trace. Git blame was designed for human authorship. It is a poor instrument for attributing a class of failure.
The Apprenticeship Ladder Is an Unfunded Liability
The third consequence sits on a longer fuse. Senior engineering judgment is largely tacit knowledge, accumulated by tracing regressions, reading strangers' diffs and being corrected. When agents absorb that work, the organization saves the quarter's labor cost and cancels the next decade's capability pipeline. This is a learning-loop argument, not a headcount argument. A team that removes productive struggle also removes the mechanism by which reviewers acquire the pattern recognition they will need to audit machine output. The firms that cut most aggressively should expect to repurchase that judgment later, at senior rates, from a market with fewer suppliers.
Attestation Is a Process Claim, Not a Correctness Proof
Regulation has its own weak point. A secure-development attestation is a declaration that specified processes were followed, not evidence that a binary is free of defects. A dependency inventory that enumerates ten thousand packages without establishing reachability can satisfy a procurement form while telling an incident commander nothing useful. Critics are right that this invites compliance theater and imposes a fixed cost that falls hardest on smaller vendors. The counterweight is equally real: voluntary assurance spent a decade failing to internalize the cost of careless shipping, and a statutory clock plus a procurement lever changes the balance sheet. Attestation raises the floor of minimum discipline. It does little to raise the ceiling of engineering quality.
What a Technical Lead Should Change This Quarter
- Split the metric. Measure review latency and change-failure rate separately from generation volume. If the first two do not improve, the agent rollout has not paid for itself.
- Keep the generation context. Attach model version, retrieval set and tool trace to every machine-authored commit, and make it retrievable during an incident rather than after one.
- Pre-authorize escalation. Publish one route a security engineer can invoke without waiting for legal to convene. A 24-hour clock does not accommodate a committee.
- Subsidize the slow work. Protect a scheduled portion of junior engineers' time for tasks an agent could finish faster. That looks like inefficiency on this quarter's ledger; it is the premium paid for having reviewers in 2029.
For non-specialists, the equivalent protection is narrower: treat an AI-generated contract, medical summary or official notice as a draft until a named human has verified it.
Six Months Out, Evidence Gets a Price
By late April 2027, expect a market split between vendors that can produce verifiable execution evidence and vendors that can only produce assurances. Procurement questionnaires will start demanding model provenance, evaluation traces and a named incident owner. Agent-native vendors will discover that per-seat pricing does not survive a customer whose bottleneck is review rather than typing. The optimistic case is that verification tooling absorbs the new demand and delivery reliability improves. The pessimistic case is that output grows, incidents grow with it, and the gap gets filled by indemnity clauses rather than engineering. Both are plausible. Only one is a strategy.