Agents buying from agents
When agents negotiate and buy from other agents, do markets stay fair and efficient, and which failure modes show up first?
Why it matters
If a business lets an agent buy or sell for it, the agent's biases, susceptibility to manipulation and bargaining habits turn into real money. Research results tell builders which guardrails to add before letting agents transact with other agents.
State of play
Newest first. Each claim is our summary of the linked primary source, with a short quote from it.
In the Bazaar auction benchmark across 11 frontier LLMs, researchers report that even the strongest pricing agent captures less than a third of hindsight-optimal profit.
However, even the strongest agent captures less than a third of hindsight-optimal profit, suggesting current LLMs are progressing in agentic commerce but leave substantial headroom.
What it does not show: A synthetic sealed-bid market with closed-form customer utilities, not live sales.
Across 9,840 LLM-to-LLM supply-chain negotiations, a preprint finds baseline models accept individually irrational contracts in 19.2% of cases, versus 0.0-0.6% for mid-tier and flagship models.
Capability also governs reliability: baseline models accept individually irrational contracts in 19.2% of cases, versus 0.0-0.6% at mid-tier and flagship, making automated profit verification the binding guardrail below that threshold.
What it does not show: One stylized bargaining game with nine models from three providers.
A preprint reports that when an LLM buyer agent shops from a verbal consumer profile, sellers can infer the buyer's willingness to pay from the dialogue alone, nearly one-for-one.
In an experiment where a language-model buyer agent shops on behalf of a verbal consumer profile, we show that seller-side inference from dialogue alone recovers willingness to pay nearly one-for-one.
What it does not show: A single experimental setup; the authors argue prompt-level fixes cannot stop this leakage.
Researchers introduce AgenticPay in a preprint, a buyer-seller negotiation benchmark, and report substantial gaps in LLM negotiation, especially long-horizon strategic reasoning.
Benchmarking state-of-the-art proprietary and open-weight LLMs reveals substantial gaps in negotiation performance and highlights challenges in long-horizon strategic reasoning, establishing AgenticPay as a foundation for studying agentic commerce and language-based market interaction.
What it does not show: Simulated markets; the abstract gives no per-model scores.
Microsoft Research says that in its marketplace tests, prompt injection redirected all payments to the manipulative agent for GPT-4o, GPTOSS-20b and Qwen3-4b.
GPT-4o, GPTOSS-20b and Qwen3-4b were very vulnerable to prompt injection: all payments were redirected to the manipulative agent under these conditions.
What it does not show: The same post reports Sonnet-4 resisted all six manipulation strategies; simulation only.
Microsoft researchers report in a preprint that in their simulated Magentic Marketplace all tested models show severe first-proposal bias, giving response speed a 10-30x advantage over quality.
Performance degrades sharply with scale, and all models exhibit severe first-proposal bias, creating 10-30x advantages for response speed over quality.
What it does not show: A simulated two-sided market, not a live one; results depend on the models and setups tested.
What would change the picture
Signals we are watching for. Our own reading of where the thread could go: not predictions, and not sourced claims.
- A live agent-to-agent marketplace publishes outcome data such as first-offer acceptance rates or manipulation incidents Would suggest: the failure modes seen in simulation can be measured in production
- A marketplace or protocol spec adds rules against first-proposal bias, such as a minimum wait or a required comparison of offers Would suggest: market designers treating agent biases as a design problem, not a model problem
- Follow-up studies rerun Magentic Marketplace or Bazaar with newer models and the first-proposal or profit gaps shrink Would suggest: the biases are model limits that improve with capability rather than structural
From the Drops
No Drop has covered this thread yet.