Writer's Agent Harness cuts AI agent costs 41% across all major models

A massive 32-author study from Writer (led by Waseem AlShikh) quantifies what they call the 'Harness Effect': the orchestration layer that assembles context, exposes tools, sequences turns, and handles governance has a bigger impact on cost than swapping models. Across 22 locked evaluation tasks and six foundation models — GPT-4o not listed but includes Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, and Palmyra X6 — they swapped only the orchestration system: a frozen conventional production loop versus Writer Agent Harness. Results were dramatic: blended cost per task fell 41% ($0.21→$0.12), median wall-clock time dropped 44% (48s→27s), and tokens per task shrank 38% (14.2k→8.8k), with task-completion quality actually improving directionally (0.78→0.81). Efficiency gains were model-invariant (33–61% cost reduction), while quality gains correlated nearly perfectly with baseline model strength (r=0.99), a phenomenon termed 'harness leverage'.

The paper formalizes token economics at the orchestration layer, including effective input pricing under prompt caching, and details six mechanism families behind the effect — from cache-shape discipline to failure-spend governance. On this workload, orchestration moved cost per task more than the entire spread of the model menu. Quality per dollar rose 82%, and task completions per million tokens jumped from 54.9 to 92.0. The authors argue that the harness is the one component whose efficiency multiplies across every model an organization runs, present and future. For enterprises grappling with ballooning agentic AI costs as token prices fall, the message is clear: stop buying capability with tokens — redesign the harness instead.