AI on Deterministic Rails

Narrative

The 2025 DORA report, drawing on nearly 5,000 practitioners, delivered the most empirically grounded verdict to date on AI in software engineering: AI functions as an amplifier of existing conditions rather than a capability creator. Teams with mature DevOps practices, well-defined workflows, and strong platform infrastructure converted AI-driven productivity gains into measurable delivery improvements; teams with fragmented tooling experienced accelerated technical debt and instability. The report's DORA AI Capabilities Model maps seven systemic capabilities — platform quality, data ecosystems, user-centricity, governance, training, communities of practice, and clear AI stance — that predict whether adoption yields net benefit. Faros AI's telemetry across 22,000 developers adds a complicating signal: individual throughput is up substantially but median PR review time has risen 441% and 31% of PRs are now merging with no review at all, a pattern they call 'Acceleration Whiplash'.

The Thoughtworks Technology Radar Vol. 33 (November 2025) documented a practitioner-level shift away from experimental AI theatre toward structured engineering. CTO Rachel Laycock noted that vibe coding had 'practically disappeared' and that the industry had moved to serious work on context, infrastructure, and security. The Radar's four themes — infrastructure orchestration for AI, the rise of agents elevated by MCP, AI coding workflows, and emerging AI antipatterns — all point in the same direction: AI is being absorbed into deterministic software infrastructure rather than displacing it. MCP, open-sourced by Anthropic in late 2024, reached near-ubiquitous vendor adoption within a year, providing a standardised integration layer between agents and deterministic systems.

The agent harness has emerged as the decisive architectural variable for agentic performance. Analysis of the Claude Code codebase established that 98.4% of the system is deterministic infrastructure — permission gates, context management, tool routing, and recovery logic — with the AI reasoning loop itself a simple while-loop. Anthropic's own engineering documentation shows that frontier models fail to build production applications without structured harness scaffolding, requiring initialiser agents, progress artefacts, and explicit testing gates. Academic work published in March 2026 confirmed that harness scaffold differences can dominate outcomes under fixed base models, establishing 'context engineering' as the successor discipline to prompt engineering.

The token-cost crisis moved from boardroom anxiety to operational emergency between late 2025 and June 2026. Stanford Digital Economy Lab research found agentic coding tasks consume 1,000x more tokens than single-turn code reasoning, with costs varying up to 30x on identical tasks. Enterprises reported hitting annual AI budgets in three months; per-developer token consumption reportedly rose 18.6x in nine months. Uber, Microsoft, and others became named examples of agentic overspend. In response, the Linux Foundation announced the Tokenomics Foundation as a FinOps-equivalent standards body, EY published a Total Cost of Agents framework, and Deloitte issued a CFO guide to AI token economics — all in spring 2026. The structural lever most consistently cited for cost control is intelligent routing: sending routine classification, extraction, and summarisation to open-weight models (DeepSeek, Qwen, Llama) while reserving frontier models for tasks where quality premium justifies the cost. One practitioner deployment report cited a 60–70% cost reduction with no measurable quality degradation using this approach.