Where AI Stands in Software Engineering
Software engineering is the most thoroughly AI-saturated profession in the economy, and it has by far the best measurement infrastructure of any of them. That combination is producing the most uncomfortable evidence base in the entire AI story. Adoption is effectively total: GitHub Copilot has crossed 20 million all-time users and 4.7 million paid subscribers with 90% of the Fortune 100 deployed; Sonar's survey of 1,100-plus enterprise engineers finds 42% of committed code is now AI-generated; Black Duck puts tool adoption at 97% across 831 enterprise engineers. What that adoption has not produced is proportionate delivery. Faros AI's two-year telemetry panel across 22,000 engineers and 4,000 teams found individual task throughput up 33.7% while review-in-progress time rose 441.5% and deployment frequency fell 11.7%. CircleCI's analysis of 28 million CI runs found feature-branch throughput up 59% year-on-year while main-branch throughput fell 7%. NBER's study of 100,000 developers found 40% more commits and 10% more releases. Three independent measurement systems, three versions of the same scissors: generation compounds, delivery does not.
Momentum is real but narrow, and it clusters where the work is bounded, verifiable, or purely investigative. Adversarial testing is the clearest advancing case: Google's Chrome security team ran a production multi-agent vulnerability-discovery workflow that surfaced a 13-year-old sandbox escape; Cisco orchestrated frontier models to scan 1.8 billion lines of code in eight weeks against an eight-year manual baseline at under 3% false positives; CISA has deployed Anthropic's Mythos model against US federal repositories. Legacy modernisation is advancing on the same logic — AI is excellent at discovery and reverse-engineering, which is why Visma's three-million-line .NET modernisation delivered a 40% effort reduction and IBM reports watsonx users growing mainframe capacity three times faster than non-users, and also why Gartner still expects more than 70% of 2026 mainframe exit projects to fail once the work moves from discovery to behavioural equivalence. Exploration and prototyping advance because throwaway code carries no verification debt. Everything that requires delegating the merge decision — autonomous approval, full autonomy, multi-agent development pipelines, agent-generated schemas — is stuck, and stuck for the same reason.
That reason is not model capability, and the industry has largely stopped pretending otherwise. Microsoft's own 16-week field study across tens of thousands of engineers is the strongest causal evidence yet in favour of agentic coding — a 24% sustained lift in merged pull requests among adopters, with a clean dose-response curve rising to 50.1% for daily users. It is also the strongest evidence of the cost: review time rose 20%, merge time 22%, and human review coverage fell from 89% to 68%. Legacy codebases saw little improvement at all. Every serious dataset now converges on the same structure: AI amplifies whatever delivery system it lands in. Plandek's analysis of 2,000-plus teams found bottom-quartile teams cut lead time roughly 50% while top performers gained only 10-15% — AI closes the gap by exposing constraints, not by removing them. The question facing engineering leaders in 2026 is no longer which assistant to buy. It is whether their verification capacity, security programme and governance architecture can absorb what they have already bought.
What's New, 2026-07-21 to 2026-08-04
The defining event of this fortnight was regulatory, not technical. EU AI Act Article 14 enforcement began on 2 August, requiring that humans overseeing AI systems actually understand the system's logic and retain override capability — the UK ICO having already established that clicking approve without comprehension does not constitute meaningful human involvement. Articles 9 and 54a took effect the same day, making continuous resilience and adversarial testing a compliance obligation for systemic-risk systems rather than a security team's preference. The timing is awkward. A VentureBeat survey of 157 enterprises published in the same window found 66% deploy or plan to deploy agents without human review, and only 5% fully trust their own evaluations. Faros data shows 31.3% of pull requests already merge with no review at all. Wiz's GhostApproval disclosure — symlink-following that lets an attacker substitute the file shown in the approval dialog across Claude Code, Amazon Q, Cursor, Google Antigravity, Augment and Windsurf — plus a confused-deputy flaw in Microsoft's Azure DevOps connector that hijacks review agents via hidden HTML in pull request descriptions, both demonstrate that even where a human is nominally in the loop, the thing they are approving may not be the thing that runs.
Nothing changed position this cycle, and that stability is itself the story: of twenty-four tracked practices, four are advancing and twenty are holding. The rest of the fortnight's evidence hardened existing diagnoses rather than reopening them. Veracode's GenAI Code Security Report across 100-plus models found the security pass rate stalled at 56%, statistically unchanged from 55% a year earlier despite two full generations of capability gains — Java sits at 30%. Secure Code Warrior and RMIT's study of 1,760 AI-generated codebases across 16 frontier models found an average of 15 confirmed vulnerabilities per codebase spanning 86 distinct weakness classes, and concluded each model carries a repeatable, predictable security fingerprint. Theori's pentest of 28 production-like AI-built applications produced 8,827 raw detections that deduplicated to 434 confirmed exploitable flaws — a useful illustration of both the real risk and the false-positive tax on finding it. Gartner's blunter framing: AI-generated code is four times more likely to contain a security defect, and the volume overwhelms application-security processes designed for a slower era. On the vendor side, a wave of verification tooling reached general availability — GitHub's Code Quality with PR-level coverage enforcement, Copilot code review with Agent Skills and MCP context injection across all paid tiers, Snyk's reachability analysis cutting false positives from over 90% to roughly 37%, Dependabot's three-day supply-chain cooldown, AWS's Agent Toolkit for infrastructure generation, Playwright's healer agent, and GitLab's agentic dependency remediation. The tooling market has correctly identified where the money now is.
Key Tensions
The bottleneck has moved downstream, and every capability gain makes it worse. This is no longer a contested claim. Faros: throughput +33.7%, review-in-progress time +441.5%, deployment frequency −11.7%, incidents-per-PR +242.7% across the adoption gradient. LinearB's 8.1 million pull requests: AI-authored PRs wait 4.6 times longer to be picked up and are accepted at 32.7% against 84.4% for human-authored code — reviewers are rationally deprioritising them. CircleCI's 28 million CI runs show the pipeline filling faster than the main branch can drain. Buying more generation capacity against a fixed review capacity is a strictly negative trade, and the organisations extracting value have understood this and funded the other side.
Security is the one curve that is not bending. Model capability has improved dramatically across two years; security pass rates have moved from 55% to 56%. Veracode's tracking of over 100 models finds reasoning models at 56% against 51% for non-reasoning — a rounding error, not a trend. Meanwhile AI-authored code carries 2.74 times the security defect density of human code and 322% more privilege-escalation paths, and IEEE's analysis of 304,362 verified AI-authored commits found over 15% introduced quality, security or maintainability issues, with 24.2% of those persisting in the codebase. Scaling has solved syntax and left semantics untouched. Anyone waiting for the next model to fix their security posture is waiting for something the data says is not coming.
Governance, not tooling, now determines whether any of this pays. Black Duck's 831-engineer survey found 97% adoption against 92% reporting governance gaps — and teams with full governance frameworks were 55% more likely to report major efficiency gains. GitLab's AI Accountability Report across 1,528 developers in six countries found 78% code faster but 79% report no change in delivery speed, with 80% having adopted tools before writing any policy. Gartner puts 88% of agentic pilots as never reaching production. The differentiator between organisations posting real returns and those posting activity metrics is not tool selection; it is whether approval gates, credential isolation, cost caps and audit trails were built before scale rather than after an incident.
An organisation's own artefacts are now part of the model's effective capability. This is the most underappreciated shift of the year. Semantic-layer grounding lifts text-to-SQL accuracy from 90.0% to 98.2% for Claude and 84.1% to 100% for GPT — while ungrounded models collapse from 91.2% on clean benchmarks to 21.3% on real enterprise warehouse schemas, and the BEAVER benchmark caps production performance at 30% even with retrieval. Stale documentation is now a runtime safety issue: a study of 145 deprecated API mappings found 70-90% context contamination with 68-100% downstream agent failure. A Thoughtworks controlled experiment refactoring a 17,155-line Rust file into 19 smaller files cut input token consumption 83%, putting a direct monetary price on code structure. Codebase hygiene, schema semantics and documentation quality have quietly become performance and cost variables, not craft preferences.
The industry is optimising to remove the human precisely as the law requires one. Approval fatigue is a documented failure sequence: agent PR volume rises, reviewers tire, the auto-approve toggle gets enabled, and OpenAI's own data shows human approval stops fall 200-fold. Anthropic flipped Claude Code's default from Auto to Manual in July citing exactly this, and Azure DevOps keeps its review agent comment-only by design — but GitHub Copilot CLI shipped Bypass Approvals and an Autopilot preview in the same period. Article 14 now makes uncomprehending approval a compliance failure in the EU, and Linus Torvalds' new Linux kernel policy points at the workable middle: not a ban, but full human responsibility for review and sign-off plus a mandatory "Assisted-by" disclosure. Provenance and accountability, rather than restriction, is where credible governance is landing.
Top 10 Evidence Items
60+ AI Coding Model Stats for 2026 (adoption-metric) — Confirms the "adoption is effectively total" baseline the whole summary is built on: 20M Copilot users, 4.7M paid seats, 90% of the Fortune 100, and 46% of deployed repo code now AI-generated. https://preuve.ai/blog/ai-coding-models-statistics-2026
Individual pace accelerated but team pace did not — analyzing public measurement data (adoption-metric) — The clearest single illustration of the "scissors" at the center of the domain: Faros AI's 22,000-engineer, two-year panel shows individual throughput up 33.7% while review-in-progress time rose 441.5% and deployment frequency fell 11.7%. https://zenn.dev/luoxi/articles/ai-team-throughput-bottleneck
Stronger with every update: How we're making Chrome and the web safer in the AI Era (case-study) — Google's own account of the production multi-agent vulnerability-discovery workflow that surfaced a 13-year-old sandbox escape — the clearest case that adversarial testing is genuinely advancing, not just being marketed as advancing. https://blog.google/security/chrome-stronger-with-every-update/
Agentic AI for Legacy Modernization: Faster and Fundable (case-study) — Documents the Visma three-million-line .NET modernization (40% effort reduction) the summary cites as proof AI excels at bounded discovery work — the advancing half of the legacy-modernization tension. https://www.simform.com/blog/agentic-ai-for-modernization/
Gartner forecast: 70%+ of 2026 mainframe exit projects will fail due to AI overestimation (industry-report) — The uncomfortable counterpart to the Visma case: Gartner's own number for why modernization stalls once work shifts from discovery to behavioral equivalence, cited directly in the summary's second paragraph. https://atmarkit.itmedia.co.jp/ait/articles/2608/03/news033.html
Microsoft's Claude Code Study: 24% More PRs, Then Cost Ceiling Hit (adoption-metric) — Covers both faces of the field study the summary calls "the strongest causal evidence yet" for agentic coding: the 24% merged-PR lift, and the 4.6x longer review delays and added vulnerabilities that came with it. https://byteiota.com/microsoft-claude-code-study-24-percent-prs-canceled/
EU AI Act Enforcement Starts August 2. Who's Governing Your Agents? (opinion) — Documents the regulatory event the summary calls "the defining event of this fortnight": Article 14's requirement that human overseers actually comprehend agent logic, landing against a 66%-no-review industry norm. https://dev.to/narko4u/eu-ai-act-enforcement-starts-august-2-whos-governing-your-agents-30f3
GhostApproval Flaw: 3 of 6 AI Coding Tools Unpatched [2026] (research-paper) — Wiz's symlink-following disclosure across Claude Code, Cursor, Amazon Q and three others is the concrete proof behind the summary's claim that the thing a human approves may not be the thing that runs — governance theater exposed as literal, not metaphorical. https://tech-insider.org/ghostapproval-ai-coding-assistant-vulnerability-2026/
Veracode Finds AI-Generated Code Security Has Barely Improved Since Last Year (industry-report) — The data behind "security is the one curve that is not bending": pass rates moved from 55% to 56% across two full generations of model capability, with reasoning models scoring no better than non-reasoning ones. https://sdtimes.com/agentic-security/veracode-finds-ai-generated-code-security-has-barely-improved-since-last-year/
Refactoring Just Got a Price Tag: Why Code Quality Is Now a Line Item on Your AI Bill (case-study) — First hard evidence for the summary's "most underappreciated shift of the year": Thoughtworks' controlled experiment shows splitting one oversized Rust file cut input token consumption 83%, converting code structure from a craft preference into a measurable line-item cost. https://reptile.haus/journal/refactoring-roi-token-cost-code-quality-ai-2026/