Enterprise AI has a bad habit: it tries to solve ROI problems with bigger context windows and more brute-force retries. A new Writer study basically tells teams to stop doing that and start optimizing the orchestration layer instead, using a “finished Writer Agent Harness” rather than changing the underlying foundation model.

The headline number that matters is the one finance people actually feel: the optimized harness drove a 41% drop in blended cost per task, from 21 cents to 12 cents, by cutting tokens per task 38%, from 14.2k to 8.8k. And the best part for anyone who has been burned by “efficiency” that silently breaks quality: task success moved from 78% to 81%, described by the researchers as directional rather than statistically significant at their sample size, while end-to-end task latency fell 44%, from 48 seconds to 27 seconds.

So what is this harness, and why does it suddenly matter as much as the model? In Writer’s framing, “AI harness” is the orchestration layer that wraps around the foundation model and turns calls into a working agent system. It routes, formats, and sequences the model’s work. The study argues this layer is not disposable glue code. It is where price of work gets set, because it controls how many rounds happen, what context gets transmitted each time, how tools are invoked, and how errors are handled. In plain terms: you can have a great model and still generate a horrible invoice if your workflow repeatedly resends bloated context or gets stuck in dead-end loops.

That connects directly to the “tokenmaxxing” problem Writer’s CTO Waseem AlShikh describes to VentureBeat. He says teams tokenmaxx because it is often the cheapest fix in the moment, and because “it’s literally how most engineers work today.” The trouble is the math is more ruthless in agentic workloads. AlShikh points out that an invoice is tokens-per-task times price-per-token, but tokens-per-task compounds when each loop iteration retransmits a growing context. Price cuts can look like progress, he says, but they can act like “an anesthetic” that hides the fact the loop itself is bleeding budget.

Writer also outlines the common failure modes of tokenmaxxing in production: simple tasks default to premium frontier models; teams use the LLM as a lazy search index by stuffing raw documents into the context window instead of retrieving exact answers; and unconstrained agentic loops can spiral when the model hits an error. The broader industry has thrown efficiency techniques at the model layer, but Writer argues those approaches tend to miss the orchestration bottlenecks. Prompt compression reduces input text, but it ignores how workflows sequence inputs across complex loops. Budgeted reasoning caps computation, but can degrade output quality when the workflow is not intelligently routed. Terse coding saves output tokens, but does not solve inefficient tool calling. Speculative decoding can speed up generation, but it does not address bloated agent architectures.

Writer’s study takes a surgical approach: hold the foundation model and the tasks constant, then change the harness. The experiments ran on six foundation models across multiple vendors and weight classes: Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, and Writer’s own model, Palmyra X6. They compared a frozen, conventional production agent loop against the finished Writer Agent Harness on 22 locked enterprise tasks covering grounding and retrieval, multi-step workflows, tool use, and content generation.

This is also where the “ROI fix” comes with a real constraint, which matters for enterprise planning and board-level risk. Smaller models like Gemini Flash 3.5 and Qwen 3.6 scored well below a usable reliability threshold on sub-agent delegation tasks (0.45 and 0.42). Sub-agent orchestration only crossed the usable reliability threshold on the two strongest models tested: Writer’s Palmyra X6 (0.86) and Claude Sonnet 4.6 (0.85). Translation: harness optimization is powerful, but it does not magically make fragile delegation reliable when the model is too small for that orchestration style. There is no free lunch, but there is still a big win when you match orchestration to model capability.

The practical part of the paper is the playbook engineering teams can actually implement. The first step is what AlShikh calls the “Two-Zone Prompt” and “Context Offloading.” The “Two-Zone Prompt” relies on prompt caching, but it only works if developers structure payloads correctly. Developers separate a “stable zone” from a “volatile zone.” Static elements like core rules, large tool schemas, and standard operating procedures go at the top of the prompt. Dynamic elements like the specific user query or recent conversational state get appended at the bottom. That ordering allows the harness to reuse the cached prefix across hundreds of calls, reducing the repeated cost of re-sending the same instructions throughout an agent’s many steps. The second lever, “Context Offloading,” targets the classic context-stuffing failure mode where the agent repeatedly sends the growing pile of history back into the model. Instead of stuffing, the harness delegates tasks like search to specialized sub-agents, giving a sub-agent only the tool and the specific query it needs, retrieving exact data, and returning a capped, clean summary to the main agent. That keeps the primary context window from filling up with raw search results.

Regulatory and procurement reality matters here too. Enterprises increasingly evaluate AI systems on reliability, cost-to-serve, and maintainability, and budgets are often locked to production performance rather than prototype demos. Writer’s framing turns the own-versus-rent question into an engineering governance issue: AlShikh says enterprises spend months on model evaluations and then “rent their orchestration off the shelf,” optimizing the smaller lever while outsourcing the bigger one. If prompt caching, tool routing, and error loops are outside your control, you can end up with unit economics that never match your production invoice. In an era where AI spend is scrutinized and policy environments demand predictable behavior, the orchestration layer becomes a strategic asset, not an implementation detail.

For founders, operators, and investors watching agentic AI roll into production, the stakes are simple: token efficiency is no longer just a cost knob. It is a system-design discipline that determines whether your margins survive real usage. Writer’s study shows the lever exists. The harness is where the work is composed, and where the price gets set. If your team is still solving ROI by feeding the model more context, you are likely paying for loops you could have engineered out.