Quick Facts
- Writer’s AI harness reduced cost per task from $0.21 to $0.12, a 41% drop, across 22 enterprise tasks tested on six foundation models.
- Median task completion time fell 44%, from 48 seconds to 27 seconds, while task quality scores held at near-parity or improved.
- Efficiency gains were model-invariant, with every model tested getting 33% to 61% cheaper, but quality gains correlated directly with baseline model strength.
Writer published research this month showing that orchestration design, not model selection, is the primary driver of token economics in enterprise AI. The findings appear in an arXiv paper titled “The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI,” dated July 2026.
The company tested its Writer Agent Harness against a conventional production agent loop across 22 locked evaluation tasks on six foundation models: Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, and Palmyra X6. Every task, prompt, model identifier, judge, and price table remained identical. Only the orchestration layer changed.
What the Numbers Show
Tokens per task dropped from 14,200 to 8,800. Cost per task fell from $0.21 to $0.12. Task quality scores moved from 0.78 to 0.81, essentially flat. Quality per dollar rose 82%. Task completions per million tokens climbed from 54.9 to 92.0.
Wall-clock time dropped from a median of 48 seconds to 27 seconds, a 44% reduction. The researchers concluded that the orchestration layer moved cost per task more than the full range of available model choices did.
How the Harness Works
The gains come from five core mechanisms: system prompt caching, interaction history compaction, tool management, retrieval strategies, and error management. The most accessible for engineering teams is what Writer calls the “Two-Zone Prompt.”
Developers split prompts into a stable zone for static elements like core rules and tool schemas, and a volatile zone for dynamic content like the user query. The stable zone sits at the top, allowing the harness to reuse a cached prefix across hundreds of calls.
CTO Waseem AlShikh described the impact: “That single separation makes prompt caching actually work and stops you from re-paying for the same instructions on every one of an agent’s thirty steps.”
The harness also delegates tasks like search to specialized sub-agents. Each sub-agent receives only the tool and query it needs, then returns a capped, clean summary. This keeps context windows narrow and reduces token waste at every step.
A strict mathematical rule governs feature additions. “If a feature adds more coordination tokens than it removes task tokens for that specific model, cut it,” AlShikh said. “Nothing in the harness is free.”
Model Size Matters for Quality, Not Cost
The research identified a pattern the authors call “harness leverage”: quality gains from the harness correlate almost perfectly with a model’s baseline capability, with a correlation coefficient of 0.99 across six models. Palmyra X6 gained the most on quality scores, followed by Claude Sonnet 4.6 and Gemini 3.1. Qwen 3.6 saw a slight quality decline.
Cost savings, by contrast, were unconditional. Every model got cheaper regardless of size or capability. A weaker model under the harness still captures its 44% to 61% cost reduction. It simply does not also get more accurate.
The Business Case
Writer, founded in 2020 and valued at $1.9 billion as of November 2024, built its platform around orchestrating complex, cross-system workflows for enterprise customers including KPMG, Intuit, Uber, and Vanguard. The company has raised $369 million total, including a $200 million Series C in November 2024.
AlShikh made the competitive stakes explicit: “Enterprises spend months on model evaluations and then rent their orchestration off the shelf, which means they’re optimizing the smaller lever and outsourcing the bigger one. Whoever owns the harness owns your unit economics, and an open framework tuned for demos is not tuned for your invoice.”
The researchers put it directly in the paper: “If the harness is the layer that composes model calls into work, it is also the layer that sets the price of work.” For software companies running agents at scale, that framing shifts where engineering resources should go.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.