AI on Deterministic Rails

Narrative

The PoC-opalypse is well-documented across independent blogs and Substacks. A May 2026 synthesis of HBR/Hyland survey data found that only 27 percent of enterprises have the connected data infrastructure needed for agentic AI, and only 17 percent had actually implemented it versus nearly half still in pilot. IBM data cited on Haverin's Substack puts only 25 percent of AI initiatives as delivering expected ROI and only 16 percent scaling enterprise-wide. The Uber case — 5,000 engineers consuming an entire annual AI budget in four months via Claude Code, with the COO admitting no clear line between token spend and delivered consumer features — became the canonical example of what several writers called token maximalism: usage divorced from outcome.

The harness-versus-model debate is the most active intellectual thread in independent technical writing between 2025 and mid-2026. Analysis of the leaked Claude Code codebase by Ken Huang and Patrick McGuinness on Substack demonstrated a three-tier permission system, a 46,000-line query engine, five-tier context compaction, and coordinator-worker multi-agent patterns whose sophistication surprised practitioners. The dominant conclusion across multiple independent writers, including Ben Dickson at TechTalks citing a UC Berkeley paper, is that system scaling has replaced model scaling as the binding constraint: once foundation models are embedded in tools and terminals, behaviour is determined by the system, not the model alone. Simon Willison, whose newsletter and blog are among the most-cited practitioner sources, identified November 2025 as the specific inflection point when Claude Opus 4.5 and GPT-5.2 made coding agents reliable enough for daily production use.

On open-weight models and cost routing, independent analysts reached a consistent conclusion by early 2026. Stanford HAI data shows the performance gap between open-weight and closed models narrowed from 8 percent to under 2 percent on Chatbot Arena by February 2025. Practitioner reporting from Particula Tech quantified a 60–70 percent cost reduction from routing 80 percent of enterprise requests to DeepSeek V4 or Qwen 3 variants, reserving frontier APIs for complex reasoning. The Digital Applied H1 2026 retrospective documented an order-of-magnitude drop in open-weight inference costs versus H2 2025, with DeepSeek V4 at approximately $0.14 per million input tokens against GPT-5 at $2.50.

Token pricing and enterprise cost shock dominate the most recent independent coverage. Goldman Sachs data, widely cited across blogs, projects a 24-fold increase in token consumption between 2026 and 2030, while Google disclosed processing 3.2 quadrillion tokens per month in May 2026, up sevenfold in one year. Simon Willison noted in an April 2026 podcast appearance that Anthropic's tokeniser changes for Opus 4.7 effectively constituted a 40 percent invisible price increase. The Veso AI and Praetorian blogs provide the most granular quantification of the corrective: deterministic scaffolding layers reduce per-query token use by 60–98 percent, depending on whether typed wrappers, schema validation, or output truncation are applied, making harness design not just an engineering choice but a direct financial control.