Patterns in the Data
Distribution of adoption
The majority of advanced engineering teams operate at L5-L8 [19][30]. Commercial investment concentrates at L4-L5 (IDE tools) and L9 (organisational tooling). No publicly documented organisation operates mature L10+ across its full engineering function. Two entities have published L11-level claims: StrongDM (3-person team, no human code review) [27] and MiniMax (self-reported 30% autonomous) [54]. Neither independently verified. The most comprehensive L10 deployments — Stripe, Ramp, Coinbase, Uber — are custom-built and not available as commercial products.
The L8 trap
Teams at L8 show +98% PR volume, +91% review time, +47% context switching, and flat DORA delivery metrics across 10,000+ developers [19]. CodeRabbit reports AI-generated PRs have 1.7x more issues [104]. The three-company RCT shows +26% individual task completion [88] while organisational delivery stays flat. The gap between individual output and organisational throughput is the central finding of the 2024-2025 productivity data.
L6 bifurcation into three sub-categories
Developer tools (Claude Code, Devin, Jules, Amp, Grok Build, Warp, Goose, Cosine) serve professional engineers. App generators (Replit, Lovable, Bolt, Base44, v0) serve broader audiences — Replit ($9B, targeting $1B ARR) reports fastest growth from non-engineers [14]. A third category — foundation models with native agentic capabilities (Kimi K2.5, DeepSeek V3.2, MiniMax M2.5) — moves orchestration from application layer to model weights, enabling agent coordination without external harnesses.
Model-native orchestration is a structural shift
Kimi K2.5 (PARL, 100 sub-agents) [48], MiniMax M2.5 (Forge, scaffold-agnostic RL) [55], DeepSeek V3.2 (thinking-in-tool-use, 1,800+ environments) [51], xAI Grok Build (8 parallel agents, Arena Mode) [8]. When orchestration moves from application to model weights, the framework boundary between L6 and L7/L8 blurs. Agents that can internally coordinate don't need external coordination infrastructure.
Open-source velocity concentrates at L6-L7 and infrastructure
OpenClaw 310K stars [17]. OpenCode 122K [11]. Gemini CLI 98K [11]. MCP 97M+ monthly SDK downloads [122]. The open-source community builds personal agents (L7) and the infrastructure they run on. Institutional patterns (L10+) remain proprietary and custom.
Chinese open-source pricing reshapes the economics
Kimi K2.5, DeepSeek V3.2, MiniMax M2.5 priced 60-90% below US frontier APIs for comparable agentic capabilities [50][52][54]. MiniMax M1 full RL training cost: $534,700 [58]. This pricing pressure affects every level from L6 upward, particularly L7 and L8 where token costs dominate.
Fintech/payments convergence at L10
Stripe, Ramp, and Coinbase independently converged on identical architectural patterns [63]. All three built custom. All three maintain human review on every merge. All three reached L10 by building on years of existing engineering infrastructure — standardised environments, comprehensive CI, documented APIs. The prerequisite for L10 is not better AI tools. It is mature engineering practice. Fabro (open-source, MIT) is the first attempt to generalise Stripe's Blueprints pattern — deterministic workflow graphs alternating with agent loops — as a reusable product [155].
Enterprise internal deployment data reveals scale
Google: 30%+ of new code AI-generated [136]. Uber: ~70% machine-generated, 95% monthly usage [117]. Shopify: AI in performance reviews [137]. Microsoft: 20-30% AI-generated [138]. Block: 5,000 weekly Goose users [96]. These represent L9-L10 operation at scale, with agents embedded in existing workflows rather than replacing them.
L10 requires pre-existing engineering maturity
Stripe's Minions built on a decade of standardised devboxes, 3 million tests, 500 MCP tools [64]. Ramp's Inspect built on existing Modal/observability stack [68]. Agent adoption compounds on existing infrastructure [64][66][68]. Organisations that lack engineering fundamentals cannot shortcut to L10 by buying AI tools.
Infrastructure layer is formalising into standards
MCP donated to Linux Foundation (Dec 2025) [123]. A2A also donated [125]. AAIF established with MCP, Goose, AGENTS.md [97]. E2B used by 88% of Fortune 100 [134]. This infrastructure is prerequisite for L10+. The standardisation pace suggests enterprise adoption will accelerate in 2026-2027.
Governance gap persists
No commercial tool provides institutional guardrails as a first-class product feature. NVIDIA NemoClaw provides runtime security but not organisational governance [106][107]. Stanford Law questioned StrongDM's accountability model [29]. Snyk reports 48% of AI-generated code contains vulnerabilities [110]. Self-healing code introduces 37.6% more critical vulnerabilities after 5 iterations [143]. The tools to build are ahead of the tools to govern.
Skill formation is a real concern
Anthropic's RCT: AI users scored 17% lower on knowledge quizzes [90]. Stanford: software developer employment aged 22-25 fell nearly 20% between 2022-2025 [144]. The Denmark study shows only 3% average time savings across 100,000 workers [92]. The NAV IT Norway study shows no measurable change in a 2-year longitudinal analysis [91]. For individual developers, the evidence for transformative productivity gains is weaker than vendor claims suggest.
Productivity measurements converge on a consistent pattern
Individual gains are measurable at L2-L6: +55% on simple completions [1], -19% to +26% on complex tasks depending on experience and methodology [3][88], and consistent gains for juniors and less experienced developers [5]. Organisational delivery metrics show no improvement at L2-L8 across all studies measuring them [19][31][32]. No controlled organisational productivity data exists for L10-L12. The hypothesis that institutional AI (L10+) delivers organisational gains remains plausible but unproven by controlled measurement.