Toolflation

Seller Agents Distort Tool Markets Through Spontaneous Generative Engine Optimization

Zhenwei Tang^1,3, Di Bai², Jintao Liu¹, Xidong Wu¹, Qianfeng Wen³, Xinwu Cheng^1
1Google    ²Google DeepMind    ³University of Toronto

Paper  |  Code

Overview

Seller agents write the listings of tools, and buyer agents read these listings when they choose which tool to call. We build tool markets from real data tools: on each of two platforms, Amazon and Instagram, ten tools from the Apify store compete for 360 questions, and every tool was run on every question's target beforehand, so the questions each tool can answer are known. In each of 20 rounds, a seller agent for each tool may rewrite the tool's description, told only to be selected more often, and buyer agents (Claude Code, Codex) choose one tool per question from the descriptions alone. Sellers run in their vendors' own harnesses: Claude Code, Codex, Grok Build, DeepSeek Harness and Antigravity. Under all five, sellers claim more and more fields, including fields that their tools return for fewer than half of the targets; we call this process toolflation. How far it goes depends on the model. Under four models, the choices stay with capable tools and buyers keep their task success. Under Gemini 3.8 Flash, sellers copy one another until the descriptions no longer tell the tools apart, buyers choose tools by their position in the list, and task success falls from 89% to about 30% on Amazon, close to random choice, and from 77% to 40% on Instagram.

Contents

This dataset holds the shelves, the capability matrices, and every description, buyer choice and seller turn of every market, from which the code reproduces every figure and table of the paper except the cost table.

Paper	Path	What it is
Sec. 3.1, Fig. 2, App. C	<platform>/tools.json, shelf.json	the ten tools, and the store search and fixed rule that chose them from 100 candidates
Sec. 3.1	<platform>/targets.json, fields.json, questions.json	30 targets, 12 fields and the 360 questions, one per target and field
Sec. 3.1, Fig. 2, App. C	<platform>/matrix.jsonl	the capability matrix: one verdict per tool, target and field (3,600 rows)
Sec. 3.4, Sec. 4, App. B	<platform>/descriptions.jsonl	every description in every round of every market, with the fields the claims judge read from it and the blanket-claim verdict
Sec. 4, Figs. 1, 3, 4, App. F, G	<platform>/choices.jsonl	every buyer choice, with the listing order the buyer saw and the arguments of its call
App. D, H	<platform>/rewrites.jsonl	every seller turn: the description before and after, each submission with its stated reason, and why the rules rejected it
Sec. 3.4, Table 1	judges.json	the second claims judge (Claude Opus 5) next to the first (Gemini 3.8 Flash)

Tools appear under the paper's codes, A0 to A9 on Amazon and I0 to I9 on Instagram, numbered by the questions they answer (A0 answers most). Every row of descriptions, choices and rewrites names its market and round:

Column	Values
buyer	claude (Claude Code with Claude Opus 5), codex (Codex with GPT-5.6 Sol)
seller	claude, codex, grok (Grok Build with Grok 4.6), deepseek (DeepSeek Harness with DeepSeek V4.1 Flash), gemini (Antigravity with Gemini 3.8 Flash)
objective	neutral in the main markets; honest or pressure in the ten-round reruns of App. F
seed	0; also 1 and 2 for the Amazon markets with Claude Code buying from Codex or Antigravity sellers
round	0 to 19 (0 to 9 in the reruns)

Markets with the same platform, buyer and seed share round 0, in which the vendors' own descriptions are used; it appears once under each of them.

Fields

tools.json: tool (code), alias (the random name agents saw, such as tool_kzqp), actor (the Apify actor), users_30d and runs (store statistics at the search), description (the vendor's description as used in round 0, with mentions of fees removed by hand), original (the vendor's unedited description), input_schema, and sample_output (the output record in the seller's product sheet). The tools are listed in shelf order, from which each buyer's listing was shuffled.

matrix.jsonl: tool, target, field and verdict, one of delivered, missing (target found, field missing), no_target or error.

descriptions.jsonl: per market, round and tool, the description buyers saw, the fields the claims judge read from it (claims), and whether Claude Opus 5 found a blanket claim of completeness (blanket).

choices.jsonl: per market, round and question, the listing order shown to the buyer, the tool it chose (choice, null if none), the arguments of that call, and any session error. The session ended at the first tool call, and whether the choice answers the question is read from the matrix.

rewrites.jsonl: per market, round and seller, the description before and after the seller's turn (after is the description of that round), and every submission with its reason and why the market rejected it (null when accepted). A row of round r is the turn taken after round r - 1.

judges.json: per platform, the number of distinct descriptions, the fields claimed under both judges for a random sample of 100 of them, and for every description of round 0 and the final round (ends).

Usage

git clone https://github.com/lilv98/Toolflation && cd Toolflation
pip install -r requirements.txt
python -m toolflation.data          # this dataset at a pinned revision, into data/
python -m toolflation.figures out/  # every figure of the paper
python -m toolflation.tables out/   # every table of the paper

from datasets import load_dataset
choices = load_dataset("lilvjosephtang/Toolflation", "choices", split="amazon")
print(choices[0]["listing"], choices[0]["choice"])

Source and personal data

The tools are public actors of the Apify store, and the round-0 descriptions, input schemas and store statistics are as their vendors published them when the shelf was searched. All later descriptions and the seller reasons were written by the seller models above, and the claims verdicts by Gemini 3.8 Flash and Claude Opus 5; use of these outputs may also be subject to the providers' terms.

Fields that can hold personal data were blanked before any tool output was stored. In the Instagram sample outputs, the names, usernames, ids and picture links of individual accounts are replaced by [redacted]. The raw tool outputs and the full histories of the agent sessions are not included, because they contain personal data of third parties.

Citation

@misc{tang2026toolflation,
  title         = {Toolflation: Seller Agents Distort Tool Markets Through Spontaneous Generative Engine Optimization},
  author        = {Zhenwei Tang and Di Bai and Jintao Liu and Xidong Wu and Qianfeng Wen and Xinwu Cheng},
  year          = {2026},
  eprint        = {ARXIV_ID},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/ARXIV_ID}
}

License

Apache License 2.0. Use of the model outputs listed above may also be subject to the respective providers' terms.

This is not an officially supported Google product. This project is not eligible for the Google Open Source Software Vulnerability Rewards Program.