Sources.

Public claims and where they come from. Figures from the study itself – the vignettes, the distributions, the token measurements – come from the interviews described in the foreword. The model, pricing and tooling layer moves monthly; treat Measurement and What Comes Next as a snapshot of July 2026.

Part I · Coding Infrastructure

GitHub Copilot (2022), Cursor (2023). Copilot entered technical preview in June 2021 and reached general availability in June 2022. github.blog/news-insights/product-news/github-copilot-is-generally-available-to-all-developers

Cursor’s agent-first workspace, routing, and Composer models. Cursor 3 introduced a distinct agent-first interface alongside the editor. Cursor’s Router selects a model per request based on task type and complexity; Composer is Cursor’s model family for coding agents. cursor.com/blog/cursor-3 · cursor.com/changelog/router · cursor.com/composer

Devin’s early-access period. Cognition introduced Devin publicly in March 2024 and initially made it available through early access. The company-specific adoption claim in the text comes from the study interviews. cognition.com/blog/introducing-devin

AI review tooling. Cursor Bugbot (agentic review, Autofix added February 2026); Anthropic's Claude Code review and /security-review; CodeRabbit, Qodo, and Graphite Agent (formerly Diamond, renamed for renewals after January 2026). github.com/anthropics/claude-code-security-review · graphite.com/blog/introducing-graphite-agent-and-pricing

Dark factories in manufacturing. FANUC's lights-out robot plants; Xiaomi's Changping facility, opened July 2024. imeche.org/news/news-article/inside-the-rise-of-unmanned-dark-factories

Part I · AI Ops

Practitioners named. Boris Cherny leads Claude Code at Anthropic (he left for Anysphere in July 2025 and returned within weeks). Peter Steinberger, creator of OpenClaw, joined OpenAI in February 2026 to work across its agent products including Codex; OpenClaw itself remains independent open source. fortune.com/2026/06/08 · techcrunch.com/2026/02/15/openclaw-creator-peter-steinberger-joins-openai

Part I · Measurement

Uber: rollout, adoption, and budget. Claude Code deployed to ~5,000 engineers in December 2025; 84% classified as agentic users by March 2026; the 2026 AI budget exhausted in four months; spend later capped at $1,500 per engineer per month per tool, with some agentic tools gated behind VP sponsorship. CTO Praveen Neppalli Naga quoted via The Information, April 2026. theinformation.com/newsletters/applied-ai · bloomberg.com/news/articles/2026-06-02 · techcrunch.com/2026/06/02

Salesforce: token spend and hiring. Marc Benioff, All-In podcast, May 2026 – Salesforce expects to spend close to $300M on Anthropic tokens in 2026 across ~15,000 engineers, alongside a continued engineering hiring freeze and a claimed 30%+ productivity gain. thenextweb.com/news/salesforce-benioff-300-million-anthropic-tokens-slack-coding · fortune.com/2026/05/28

List pricing and the cost arithmetic. Anthropic published rates: Opus $5 / $25 per MTok, cache write $6.25 (5-minute) or $10 (1-hour), cache read $0.50; Fable $10 / $50, cache read $1.00. The $1K, $10–12K and $20–23K figures in the text are cache-adjusted, applying the July 2026 session mix quoted there – roughly 94% cache reads; naive tokens-times-output-rate math overstates them by more than 20×. platform.claude.com/docs/en/about-claude/pricing

Subscription tiers. Claude Max 20x and ChatGPT Pro 20x at $200/month. OpenAI added a $100/month 5x Pro tier in April 2026. claude.com/pricing · techcrunch.com/2026/04/09/chatgpt-pro-plan-100-month-codex

Lower-cost substitution. GLM-5.2 released June 2026 (open weights, long-horizon agentic coding, low-cost plans). Cursor's Composer 2.5 is built on Moonshot's Kimi K2.5; the ~10–60× cost-per-task advantage is Artificial Analysis's independent measurement. venturebeat.com/technology/z-ais-open-weights-glm-5-2 · artificialanalysis.ai/articles/cursor-composer-2-5-coding-agent-index

Concentration of AI output: 46×, Gini 0.77. Cursor, 2026 Developer Habits Report (data through May 2026): p99 ships 46× the median's AI-written lines and 15× merged PRs; p90 ships 10× and 4×. cursor.com/insights

Top 5% of teams doubled throughput; median +4%. CircleCI, 2026 State of Software Delivery, February 2026, from 28M workflows. Throughput measured as daily workflow runs. circleci.com/blog/five-takeaways-2026-software-delivery-report

Complexity scoring in the open. Lemonade's complexity-analyzer, an open-source CLI that uses an LLM to score pull-request difficulty. github.com/lemonade-hq/complexity-analyzer

Merge-readiness. Cognition’s FrontierCode evaluates whether repository maintainers would merge model-generated pull requests, including correctness, regression safety, tests, scope, style, and repository fit. cognition.com/frontiercode

DX: AI adoption and pull-request throughput. A longitudinal study of more than 400 companies from November 2024 through February 2026 found AI adoption up 65%, median PR throughput up 7.76%, and mean throughput up 13.1%; most companies landed between 5% and 15%. getdx.com/news/new-data-ais-impact-on-engineering-velocity-is-more-modest-than-expected

Engineering-analytics vendors. DX, Jellyfish AI Impact, LinearB, Hivel and Worklytics all now report on AI-assistant telemetry. DX also publishes an AI measurement framework connecting adoption, system performance, developer experience, and business impact. getdx.com/blog/ai-roi-engineering · jellyfish.co/blog/measure-ai-impact-copilot-cursor-gemini-sourcegraph · linearb.io/use-case/measure-ai-impact

Part I · Security & Compliance

EchoLeak (CVE-2025-32711). Indirect prompt injection in Microsoft 365 Copilot allowing information disclosure over a network; Microsoft CVSS 9.3. nvd.nist.gov/vuln/detail/CVE-2025-32711

Shai-Hulud. Self-replicating npm worm, September 2025, spreading via stolen publish tokens; the November 2025 second wave reached ~796 packages. krebsonsecurity.com/2025/09/self-replicating-worm-hits-180-software-packages · securitylabs.datadoghq.com/articles/shai-hulud-2.0-npm-worm

SANDWORM_MODE. Disclosed by Socket's threat research team, February 2026; corroborated by CrowdStrike and Endor Labs. Named for the malware's own SANDWORM_* switches, not the Russian APT. socket.dev/blog/sandworm-mode-npm-worm-ai-toolchain-poisoning

Spotlighting. Hines et al. (Microsoft), “Defending Against Indirect Prompt Injection Attacks With Spotlighting” – attack success from over 50% to below 2% across GPT-family models. arxiv.org/abs/2403.14720

The dual-LLM / quarantine pattern. Proposed by Simon Willison, April 2023; credited in OWASP's prompt-injection prevention cheat sheet. simonwillison.net/2025/Jun/13/prompt-injection-design-patterns

AI-gateway and guardrail vendors. Noma and Lasso remain independent; Lakera was acquired by Check Point (November 2025) and Prompt Security by SentinelOne (September 2025). MintMCP provides MCP-specific gateways. checkpoint.com/press-releases/check-point-acquires-lakera · sentinelone.com/blog/sentinelone-acquires-prompt-security

Adversary use of AI: 89% year over year. CrowdStrike, 2026 Global Threat Report, February 2026. crowdstrike.com/en-us/press-releases/2026-crowdstrike-global-threat-report

AI performing 80–90% of a campaign. Anthropic, “Disrupting the first reported AI-orchestrated cyber espionage campaign,” November 2025. anthropic.com/news/disrupting-AI-espionage

AI red teaming. FireCompass (continuous automated red teaming); Zscaler (via its SPLX acquisition, announced November 2025); HackerOne (researcher-led AI red teaming); Obsidian Security (AI security posture management). firecompass.com/continuous-automated-red-teaming · zscaler.com/press

Retention, training, and zero data retention. OpenAI retains API abuse-monitoring logs up to 30 days by default, with ZDR requiring prior approval; the Responses API stores application state by default unless store: false is set. Anthropic deletes inputs and outputs within 30 days on the backend, with longer retention for policy enforcement and legal hold. developers.openai.com/api/docs/guides/your-data · privacy.claude.com

“Weights and alpha.” Alex Karp, CNBC, July 2026, alongside Palantir's accompanying statement on model dependence. cnbc.com/2026/07/01/palantir-karp-open-ai-anthropic-tokens.html

Pass-through routing. Together AI's privacy documentation: pass-through models forward prompts and responses to the upstream provider under that provider's policy, and prompt storage must be enabled to use them. docs.together.ai/docs/privacy-and-security

Qwen tiering. Alibaba previewed Qwen3.8-Max in July 2026 alongside its personal Token Plan; open weights announced but not released at the time of writing. marktechpost.com/2026/07/19

Political bias by model origin and prompt language. Lim & Röttger, “Bias in the East, Bias in the West,” Findings of EACL 2026 – 36,000 parallel prompts across 60 political issues. aclanthology.org/2026.findings-eacl.122

Suppression in DeepSeek. Carragher, Williams & Carley (Carnegie Mellon), “Information Suppression in Large Language Models” – 646 sensitive prompts, comparing chain-of-thought with final output. arxiv.org/abs/2506.12349

Local and distributed inference. Apple, WWDC26 sessions 232 (“Run local agentic AI on the Mac using MLX”) and 233 (“Explore distributed inference and training with MLX”), citing a 1.6-trillion-parameter DeepSeek model requiring more than 800GB sharded across several Macs. NVIDIA's Nemotron family publishes models, weights, datasets, and training recipes for on-premises or private-cloud use. developer.apple.com/videos/play/wwdc2026/232 · developer.nvidia.com/nemotron

The compliance stack. SOC 2 criterion CC8.1 quoted from the AICPA Trust Services Criteria. AIUC-1 is positioned as the first agent-specific standard, using independent audits with at least quarterly technical testing; ISO/IEC 42001 covers AI management systems and the NIST AI Risk Management Framework provides voluntary risk guidance. aiuc-1.com

Part I · How the Strongest Teams Build It

Internal agents. Ramp's Inspect runs on open-source OpenCode inside Modal sandboxes, writes over half of merged PRs, and is now >80% self-written. Stripe's Minions ship ~1,300 PRs a week on a forked Goose. Shopify open-sourced Roast; Coinbase runs Forge (formerly Claudebot, then Cloudbot). modal.com/blog/how-ramp-built-a-full-context-background-coding-agent-on-modal · blog.bytebytego.com/p/how-stripes-minions-ship-1300-prs · github.com/shopify/roast

The buy-side alternative. Factory's Droids – a commercial agent runtime with model routing, integrations, permissions, and observability. factory.ai/product/droids

Goose under the Linux Foundation. Block contributed goose to the Agentic AI Foundation, formed December 2025 alongside MCP and AGENTS.md; the move completed in April 2026. linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation

Part I · What Comes Next

The collaboration stack. Cursor's Origin, announced June 2026 as a “Git forge for the agentic era,” waitlist-only at the time of writing. Entire, founded by former GitHub CEO Thomas Dohmke, previewed its distributed Git network on July 2026. Block's Buzz, an open-source agent workspace built on signed events, launched July 2026. BAND provides shared rooms, memory, and identity for agents across frameworks. cursor.com/origin · geekwire.com/2026 · siliconangle.com/2026/07/21 · venturebeat.com/orchestration

Part II · Convert the People

Legacy-modernization plays. The Strangler Fig, Branch by Abstraction, Parallel Change, event interception, and transitional architecture all replace a system through temporary seams while old and new implementations coexist. docs.aws.amazon.com/prescriptive-guidance/latest/cloud-design-patterns/strangler-fig.html · martinfowler.com/bliki/BranchByAbstraction.html · martinfowler.com/bliki/ParallelChange.html · martinfowler.com/articles/patterns-legacy-displacement

AI-assisted modernization. Microsoft’s GitHub Copilot App Modernization follows an assess–plan–execute workflow; Meta’s large-scale ingestion migration used shadow jobs, automated validation, and staged promotion before cutover. learn.microsoft.com/en-us/dotnet/core/porting/github-copilot-app-modernization/overview · engineering.fb.com/2026/05/12/data-infrastructure/migrating-data-ingestion-systems-at-meta-scale

Forward-deployed engineering ventures. Anthropic's Ode, a joint venture with Blackstone, Hellman & Friedman and Goldman Sachs, launched May 2026; OpenAI's majority-owned Deployment Company, announced May 2026. techcrunch.com/2026/07/15 · forbes.com/sites/janakirammsv/2026/05/28

Comprehension debt. Term coined by Addy Osmani, late 2025; republished by O'Reilly Radar. addyosmani.com/blog/comprehension-debt

“I don't prompt Claude anymore.” Boris Cherny, in conversation, June 2026. lucumr.pocoo.org/2026/6/23/the-coming-loop

Vibe coding and agentic engineering. Both terms are Andrej Karpathy's: “vibe coding” from February 2025 (Collins' word of the year for 2025), and “agentic engineering” proposed in February 2026. x.com/karpathy/status/2019137879310836075

Part III · Assemble the Community

Model bias toward familiar technology. Twist, Harman, Syme, Noppen, Yannakoudakis, Nauck & Zhang, A Study of LLMs’ Preferences for Libraries and Programming Languages, Findings of ACL 2026. Across eight models the pull toward the popular is universal: the same three Python libraries top every ranking, and Python stays the dominant choice in 58% of tasks where it is the wrong tool. arxiv.org/abs/2503.17181

Roles blending; the whole organisation moving at one speed; the 41× benchmark. All three from Cat Wu, Head of Product for Claude Code, “Product management on the AI exponential,” March 2026. The 41× figure is Anthropic's internal task-length benchmark, measured from Claude Sonnet 3.5 to Opus 4.6. claude.com/blog/product-management-on-the-ai-exponential

The 1:6 to 1:10 PM ratio. Long-standing industry rule of thumb, traceable to Marty Cagan and consistent with subsequent practitioner surveys. bringthedonuts.com/newsletter/ideal-ratio-engineers-to-product-managers

APIs as the interface between teams. The 2002 Bezos mandate at Amazon, described publicly in Steve Yegge's 2011 platform post. nordicapis.com/the-bezos-api-mandate-amazons-manifesto-for-externalization

Pairing engineers with domain experts. Uber's Agentic Pods programme, described publicly by CTO Praveen Neppalli Naga, mid-2026: ~30 engineers, two-week pods, sixteen functions in two months. thestateofai.com/news/uber-unveils-agentic-pods-structure

The pre-turn productivity measurement. METR (Model Evaluation & Threat Research), randomized controlled trial, July 2025: 16 experienced open-source developers were 19% slower on their own repositories with early-2025 AI tools, while estimating they had been 20% faster. A late-2025 re-run of the same cohort reversed the direction, with the authors cautioning it is a weaker signal – in part because participants increasingly declined to work without AI. metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study