A Harness Reducing over-Defensive Writing in Agents by Isolating Objections

Xiaochen ChangarXiv 2026 · arXiv

TL;DR

AI agents often clutter reports with "over-defensive" writing—anticipating and disclaiming criticisms that weren't even raised. This paper introduces a multi-agent "harness" that isolates the review process from the writing process, reducing defensive clutter by over 96% while significantly improving the retention of original source claims.

What Problem Does This Paper Address?

When we ask an AI agent to expand a draft into a formal report, it often suffers from over-defensive writing. This occurs when the model introduces a claim stronger than what is in the source material, only to immediately walk it back or add unnecessary qualifications.

For example, if a source draft mentions reducing processing time, an over-defensive agent might add a sentence stating that "delivering technology alone does not guarantee project success." While technically true, this qualification wasn't requested and distracts the reader from the actual goal.

The authors hypothesize that this happens because agents are essentially writing for an "adversarial AI reviewer" rather than a human reader. Because the model "thinks" about potential criticisms while drafting, those criticisms leak into the final text—a phenomenon known as contextual entrainment. Simply telling an agent "don't be defensive" often fails because the objection is already present in its context window, making it hard to ignore.

A comparison showing how a simple expansion introduces unnecessary defensive qualifications compared to the source.

Methods

The researchers developed a three-stage "harness" that separates the roles of reviewing, filtering, and writing into independent sub-agents to ensure context isolation.

  1. Reviewer Agent: Analyzes the source files and the task to generate "atomic" technical review points—specific questions or concerns that a careful reader might have.
  2. Filter Agent: Estimates the probability of a reader raising these concerns. It selects only the most relevant points based on a threshold, converting them into affirmative writing objectives (e.g., "Explain how X works" rather than "Defend against the idea that X fails").
  3. Writer Agent: This agent starts with a fresh context. It receives the source material and the filtered list of objectives but is strictly forbidden from conducting its own defensive review.

Crucially, the writer is given a "material-strength constraint": it must not invent new facts to answer a reviewer's concern, ensuring that any additions remain grounded in the original evidence.

Datasets

The study used a dataset of 25 business-project case drafts. These were short English drafts containing facts, goals, and implementation details. The task for the agents was to expand these into formal, ~500-word reports.

The evaluation compared three methods:

  • Direct Expansion: The baseline (no special instructions).
  • Prohibition: Using prompts like "Avoid defensive writing."
  • The Harness: The proposed multi-agent workflow.

The models used were GPT-6.1 Sol for the harness agents and DeepSeek-V4.1-Flash-1017 for assessing the outputs.

Key Insights and Ideas

  • Context Isolation is Key: The most effective way to stop an agent from being defensive is to prevent it from "seeing" the objections in the first place. By the time the Writer Agent starts, the messy process of brainstorming criticisms has been filtered and abstracted away.
  • Atomic vs. Coarse Review: Breaking down criticisms into single, concrete points (atomic) is far more effective than providing broad, compound feedback. Atomic points led to a 98% uptake by the writer, compared to only 89% for coarse points.
  • The "Entrainment" Trap: The authors found that if a concern has even a 10% probability of being relevant, an agent will often write it into the output if it's present in the context. This confirms that LLMs struggle to ignore "low-probability" thoughts once they are "on their mind."
Table showing that atomic review points and specific instructions against weakening claims lead to better retention of source material.

Key Results

The harness outperformed both direct expansion and simple prompt-based prohibitions across all metrics:

  • Defensive Reduction: Defensive propositions dropped from 59 (in direct expansion) to just 2—a 96.61% reduction.
  • Fact Integrity: While simple prohibition prompts led to 51 "unsupported factual additions" (hallucinations), the harness produced zero.
  • Claim Retention: The harness retained 93.12% of the original source claims, compared to only 80.35% for direct expansion.
Table 5: Defensive Writing Across 25 Cases

Limitations and Future Work

While the results are striking, the authors acknowledge several areas for further study:

  • Scale: The evaluation was limited to 25 cases. A larger, more diverse test set is needed to confirm these findings across different domains.
  • Human Validation: The assessment was performed by an LLM. While LLM-as-a-judge is a common practice, human annotation would provide a more nuanced look at "necessity" vs. "over-defense."
  • Review Robustness: The researchers noted that two different reviews of the same text only overlap by about 60%, suggesting that the initial "Reviewer Agent" stage could be made more consistent and comprehensive.

Summary updated

Abstract

When expanding a report, writing with agents may introduce a stronger claim than the article makes and then disclaim it. This is called over-defensive writing. We assume that this is because the agent is writing for responding to an adversarial AI reviewer, instead of a human reader. The defensive context may also influnce the consequent when we simply ban defensive writing in the output through prompts. In order to reduce over-defensive writing, we produce a harness that separates the reviewer, filter, and writer. The writer agent receives only the problems choosen by the filter, which is an independent context. We evaluate this harness on 25 business-project drafts expanded into formal reports, the result shows that the defensive propositions decreased from 59 to 2, which is a 96.61\% reduction, no unsupported factual additions were identified while the source-claim retention rises from 80.35% to 93.12%.
BibTeX
@misc{chang2026harness,
  author = {Xiaochen Chang},
  title = {{A Harness Reducing over-Defensive Writing in Agents by Isolating Objections}},
  year = {2026},
  eprint = {2610.36717},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.36717},
}
Figure