The Next Fifteen Years

A forecast built from first principles
Section future / 03-domains / cognitive / software.md

Software Engineering#


Contents

Most-affected large profession. Not because coding disappears, but because the unit of work changes from writing code to specifying, reviewing, and integrating.

The numbers that matter#

Expect 2–4× throughput on greenfield and near-zero gain on legacy integration. The ratio between those is where the real story is - it means the benefit is concentrated in new work and in organizations without accumulated context debt, which is very nearly the opposite of where most engineering labor currently sits.

Total developer employment probably rises through 2030 (Jevons on software demand), while composition shifts hard toward senior/architectural and away from junior/implementation.

Work typeAI leverageEmployment pressure
Greenfield features, boilerplate, testsHighJunior implementation ↓
Legacy integration, partial rewritesLow–mediumSticky; context is the moat
Architecture, incident judgment, multi-team designLow (verification expensive)Senior demand ↑
Spec / review / eval harnessesMedium–highNew hybrid roles

The bottleneck migrates to review#

Amdahl's law applies to the team, not just the machine: if generation speeds up 5× and review does not, review becomes the schedule. This is already the observed shape in AI-heavy teams - pull requests queue at the senior engineers whose approval makes the code real, and the organization's effective throughput is the throughput of its trust, not of its typing. Three consequences follow. First, the economic premium moves to whatever makes verification cheap - test coverage, typed interfaces, deterministic builds, observability - so codebases built to be checked compound their advantage over codebases that must be understood. Second, the pressure to automate review itself is enormous and partially self-defeating: model-reviewed model code collapses the independence that made review a control (Uncertainty 5 in miniature). Third, "senior demand ↑" in the table above is really reviewer demand, and reviewing is a skill trained by doing the work that juniors no longer do - the apprenticeship gap eating its own antidote.

The open-source commons shows the failure mode early and in public. Maintainers are a review bottleneck with no budget, and the flood arrived first as noise: the curl project publicly documented a wave of AI-generated bogus vulnerability reports through its bug bounty (2024, per maintainer Daniel Stenberg), each costing scarce expert hours to refute. Free generation plus expensive verification is a tax levied by the many on the few, and volunteer infrastructure pays it first. If the commons responds by closing - reputation-gated contribution, paid triage - the open-by-default era of software ends not by license change but by review economics.

Junior employment as the leading edge of Game 4#

Software is where Game 4 is measurable first, because the junior task set (boilerplate, tests, first-draft PRs) is exactly the high-leverage greenfield row in the table, and because posting data is liquid. The B1 baseline already shows the leading edge; the software-specific claim is that stabilization of junior software postings while other knowledge professions keep falling would mean the inversion path in Uncertainty 3 is live here first - dense AI feedback compressing the path to competence where verification is cheapest. Continued decline in junior software postings through 2029, even in a hiring recovery, is the commons-failure path at its purest.

Review is the scarce seat. Amdahl on the team means senior PR capacity sets throughput. Tools that flood generation without shrinking review time raise queue length, not shipped value - the open-source bogus-report tax scaled to every company that hires juniors as "AI amplifiers" without growing reviewer headcount.

What it does to SaaS#

Software's marginal cost falls toward zero, threatening the SaaS model - per-seat pricing, high margin, defended by switching cost.

If a bespoke internal tool costs $5k instead of $500k, the long tail of vertical SaaS gets eaten from below. The vendors that survive hold something other than the software: proprietary data, a compliance position, a network, or distribution. → Game 3

Seat → outcome (timeline consistency)#

2026–2028 and B5: the commercial signature of agentic reliability is pricing on outcomes, not seats. Software is the first large market where that shift is visible.

PricingAssumesBreaks when
Per-seatHuman operator per licenseAgents do the work; headcount ≠ value
Usage / tokensMetered cognitionRace to zero on inference (inference economics)
Outcome / success feeAttributable results + liabilityReliability and indemnity unclear

Vendors that cannot reprice watch ARR per customer fall while usage rises. Buyers that accept outcome pricing reveal belief that unsupervised work is real - more honest than benchmarks.

The apprenticeship problem is sharpest here#

This domain is the leading indicator for Game 4 because the exposure is cleanest: AI is best at exactly the work that used to be how people learned. Boilerplate, first-pass implementation, test writing, and bug triage were never economically valuable in themselves - they were the tuition.

Removing them is efficient for every individual team and quietly catastrophic for the pipeline.

Aligned with law (same pyramid economics) and education (credentials + missing junior years). Mid-2026 data already shows junior software postings and entry-level tech hire shares collapsing - see Game 4 tables.

Discriminating test (shared with labor): if junior eng hiring fails to recover when aggregate tech hiring does (~2027–28), substitution share was large. → B1

Uncertainty 3 here first#

Uncertainty 3: AI as dense mentor could shorten novice→expert if feedback is grounded (tests, types, prod metrics). Software has the cheapest ground truth for that inversion - science-like verification on code. If the inversion fails here, it is unlikely to save law or consulting.

Security and supply chain#

Cybersecurity: model-written code at volume without review expands the vuln surface; autonomous fix loops help only if verification holds. SBOMs, signing, and least-privilege tools become inelastic complements to cheap generation.

What to watch#

SignalReading
Entry-level eng posting shareApprenticeship gap
Seat vs outcome revenue mix at major AI-devtool vendorsAgentic commercial threshold
Vertical SaaS churn / internal-tool build ratesLong-tail SaaS eaten
Time-to-merge / incident rates with AI-heavy PRsThroughput vs quality
Intern / new-grad conversion rates at large techPipeline health

Failure modes#

Context debt is the moat that generation cannot buy. Greenfield throughput without integration into decades of partial systems is demo economics. Score production PRs merged into brownfield repos and incident rates after agent merges - not greenfield benchmark suites alone.


Related: Game 4 · Game 3 · Inference economics · Startups (formation and selection when shipping is cheap) · Law · Cybersecurity · 2026–2028 · B1 · B5

View markdown source

select · Enter open · Esc close