The program wrapped around an AI model, its system prompt and its tools, barely changes how many tasks an agent solves but can change what the work costs by up to 2.6 times on the same model and task, according to a paper posted on arXiv on 9 November by researchers led from UC Berkeley, with co-authors at Northwestern University and the University of Illinois.
In the paper’s own test, Anthropic’s Opus 5.5 cost about $0.42 per task inside Claude Code against about $0.28 in a bare harness with only a bash tool, while solving about the same share of tasks, about 89.5% against about 90%, by our reading of the paper’s Figure 1. That gap is about $1,400 per 10,000 tasks, by our arithmetic.
Business Pill · Tokens Explained
An evergreen Business Engineer explainer on tokens, the unit AI reads, writes and bills in. It teaches the general idea only and does not discuss this paper.
The key insight: As we read the paper, the harness sets the bill more than the score: on the same model and tasks, the bare bash-only harness solved about as many tasks as Claude Code while Opus 5.5 cost about $0.28 per task against about $0.42, by our reading of its Figure 1.
What the Paper Tested
The authors ran 100 tasks, 20 from Terminal-Bench 2, 40 from SWE-Bench and 40 from SWE-Bench Pro, on five models through four harnesses: Claude Code and OpenAI’s Codex, installed unmodified, and two simple harnesses built to isolate the harness itself. Basic-full has a generic software engineering prompt and six tools; Basic-bash has no system prompt and only a bash tool.
The paper reports that “Even the bare-bones Basic-bash harness provides task accuracy similar to that of Claude Code and Codex, with even better results on some models”. It says the newer models “are smart enough to navigate through the tasks without the generic guidance in the system prompts”.
Cost moved more. By the paper’s Table 2, Claude Code starts each session with 4.1K tokens of system prompt and 8.7K tokens of tool definitions for 18 tools, or 12.8K tokens in all by our arithmetic, against 0.24K for the bash-only harness, about 53 times as much. Looking at the first model call, the authors found many of those instructions are “for user experience, such as how to present data through the terminal or how to refer to the user when their pronouns are unknown”.

Where the Money Goes
The paper finds that re-reading context already in the conversation is the largest cost, even with caching, taking up to 73% of the total. Tool calls caused up to a 26 times cost increase at the same token count, and the agent used only 27% of tool outputs on average, the authors write.
File searches took about 43% of agent turns and 28% of cost across the benchmark runs, the paper says. Cache reads are paid using the full input context on every call, the authors explain, so both longer context and extra turns add to the bill.
What Agency Does
The authors built Agency, which they describe as “a harness interposition platform inspired by Valgrind”. It sits between an off-the-shelf harness such as Claude Code or Codex and its environment, and changes context and tool calls “without any modification to the harness binary or the LLM backend”.
On it they ran five techniques, including compacting tool outputs, removing stale context and building a searchable index of the files. Combined, cost fell to 0.77 of the baseline on average, a 23% cut by our arithmetic, while the accuracy ratio was 0.97, the paper reports. Claude Code saved least, at 0.81 of its baseline; the abstract cites cost reductions of up to 35%.
How It Compares
The finding lines up with Garry Tan’s open harness test: our piece on that test found the costliest harness on GPT-6.1 Sol spent 2.18 times as much per solved task as the cheapest. Our HKU and Salesforce piece covers a paper that trains the agent and its harness together, and our Amazon multi-agent piece covers a setup that used 4.7 times the tokens for an 8-point gain.
The paper cites earlier work, HarnessTax by Pan and co-authors including Ion Stoica, for the finding that the harness matters little for accuracy on frontier models. Stoica is also an author of this paper.
The Structural Read
As we read the cost breakdown, the expense sits in repetition: the paper says cache reads are paid using the full input context on every call, so 12.8K tokens of fixed instructions, by our arithmetic from its Table 2, are paid for again on every turn.
As we read the Agency results, the savings come without touching the model: the paper reports cost at 0.77 of the baseline with an accuracy ratio of 0.97, by changing only what the harness sends.
Jongseok Park, Ion Stoica, Alvin Cheung and co-authors, arXiv 2611.12122, 9 November 2026
“We conclude that harness designs can significantly affect agent execution cost while minimally affecting task accuracy.”
Three Implications
FOR TEAMS RUNNING AGENTS AT VOLUME On Opus 5.5 the gap between Claude Code and a bash-only harness is about $1,400 per 10,000 tasks, by our reading of the paper’s Figure 1 and our arithmetic.
FOR HARNESS MAKERS The paper says many of the extra instructions in Claude Code and Codex are for user experience, such as how to present data through the terminal.
FOR MODEL BUYERS The paper reports that the newer models chain shell commands or write scripts to reach their goal without special tools.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Agentic Harness War.
The framework argues: “The surface, not the model, governs autonomy.”
As we read it, this paper adds the price of that surface: on the same model and tasks, the harness moved cost per task far more than accuracy. The framework predates the paper and does not discuss it.
What Is Not Established
The per-task dollar and accuracy figures here are our reading of the paper’s Figure 1, which plots them without data labels, so they are approximate, to about a cent and a point. Its prices include a 10% AWS Bedrock surcharge, and the authors note that their benchmark tasks are 2 to 3 times shorter than real-world traces.
The paper is a preprint. In the body, the authors put the harness effect on cost at 2 to 3 times; by our reading of Figure 1, the widest ratio is on GPT-6.1 Sol, about $0.26 against about $0.09 per task. Neither the authors nor Anthropic or OpenAI were contacted for this piece.
The Bottom Line
A Berkeley-led paper finds that the harness around an AI agent changes cost far more than accuracy, by up to 2.6 times on the same model and task, its abstract says. On Opus 5.5, Claude Code cost about $0.42 per task against about $0.28 for a bash-only harness at a similar pass rate, by our reading of its chart.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Every finding, table figure and quote comes from the paper “System Optimizations for Cost-Efficient Agent Harnesses” by Jongseok Park, Ion Stoica, Alvin Cheung and co-authors, posted on arXiv (2611.12122) on 9 November 2026 and read in full on 16 November; we found it through DAIR.AI’s weekly list of top AI papers. Per-task costs and accuracy are our approximate reading of the paper’s Figure 1; token totals, gaps per 10,000 tasks and percentage cuts are our arithmetic.
Sources: arXiv 2611.12122, System Optimizations for Cost-Efficient Agent Harnesses (9 Nov 2026) · the paper as PDF (Figure 1, Tables 1-2) · DAIR.AI, Top AI Papers of the Week









