LLM agents increasingly interact with other agents on behalf of users, yet how users in the loop affect their agents' behavior--for better or worse--remains poorly understood. We study this question through the lens of negotiation, where agents must represent users while strategically bargaining with other agents. To isolate the effect of user presence, we develop PRESENT, a controlled evaluation framework that compares negotiation settings with and without an interactive user, while holding the game, objectives, and user preferences fixed across conditions. Across 11.2k runs spanning 7 games and 8 models, we find that user presence significantly changes both outcomes and behavior. Specifically, negotiation parties with a user obtain higher payouts (+13.7% on average) and their agents negotiate more aggressively. Surprisingly, user presence also induces misaligned behavior, both toward counterparties (where its rate roughly doubles) and toward users (where it emerges in 14.3% of games). These misaligned behaviors are uncorrelated with outcomes and persist even when the user sends only templated, neutral messages. These findings expose two blind spots in current evaluation practice: static instructions miss behaviors that emerge only when users are present, and outcome-level metrics cannot detect them.
Figures & tables
Figure 1: Users change how agents negotiate. While most negotiation benchmarks do not include users in the loop, our evaluation framework compares otherwise identical negotiations with and without an interactive user to isolate the effect of user presence (top). We find that users shift agent behavior in three ways (bottom). Agents with users obtain higher payouts and negotiate more aggressively on average; however, they also display more misaligned behavior, toward both their counterparties and their own users.
Figure 2: Overview of Present . We model bilateral negotiation with an interactive user who can provide initial guidance (A), be consulted during negotiation (B), and must consent to the final deal (C). This common interface allows us to measure the effects of user presence consistently across a range of different negotiation games. We provide an example of a full negotiation transcript in § A.9 .
Game
Negotiation terms
Source
Buy-Sell
A seller and a buyer negotiate a price P for one good.
Negotiation Arena Bianchi et al. (2024)
Ultimatum
Two players negotiate a split of 100.NistheamountofmoneyallocatedtoPlayer2byproposalx$ .
CaSiNo
Two campers divide a fixed stock of food, water, and firewood. For camper i and item type j , qij is the quantity and vij is the camper’s value per unit.
CaSiNo Chawla et al. (2021)
Deal or No Deal
Two players divide a pool of books, hats, and balls. For player i and item type j , qij is the quantity and vij is the player’s value per unit.
Deal or No Deal Lewis et al. (2017)
Service Contract
An engineer and a manager negotiate the hourly rate R and the contract duration.
PrincipalBench Li and Shi (2026)
Car Sale
A seller and a buyer negotiate the price P of a used car.
Table 1: Negotiation games, adapted from four sources and covering both single- and multi-issue negotiations.
Behavior
Definition
Example
User-facing
Misreporting
Agents report to the user information that is inconsistent with the agent–agent history, or omit an important alternative.
The agent reports to the user that the counterparty offers 52,buttheactualofferis48.
Hallucinated action
Agents falsely claim they did an action.
The agent claims that it has sent a counteroffer without actually sending it.
User-facing unsupported urgency
Agents pressure their users with unsupported claims of urgency.
The agent tells its user that it will finalize the deal unless the user objects within two minutes, though the game has no time limit.
Counterparty-facing
Misunderstanding
Agents communicate inaccurate information about the negotiation state in messages to the counterparty.
The agent says that the counterparty increased its offer by 1,000,althoughtheofferroseonlyfrom10,800 to $11,000.
Table 2: Misaligned-behavior categories. We group by whether the behavior is directed at the agent’s user or at the counterparty. Examples are paraphrases of actual logs; more verbatim examples can be found in § A.10 .
Game
Claude
DeepSeek
GPT-5.6
Qwen
Pooled
(95% CI)
Haiku 4.5
Sonnet 5
V4 Flash
Luna
Terra
3.6 Plus
3.7 Plus
3.8 Flash
Buy-Sell
43.5 *
25.6 *
41.5 *
43.8 *
24.9 *
35.1 *
38.2 *
33.7 *
35.8 ± 6.3 *
Ultimatum
23.0 *
17.0 *
31.0 *
31.7 *
20.8 *
23.5 *
28.1 *
32.6 *
25.9 ± 5.3 *
CaSiNo
2.7
− 5.3 *
0.0
3.5
1.0
1.6
0.2
5.2 *
1.0 ± 2.9
Deal or No Deal
1.3
− 2.9 *
− 5.2 *
4.4 *
− 1.0
0.7
− 3.7 *
1.5
− 0.7 ± 2.6
Table 3: Payout change of the user-side agent in With User relative to No User . Cell color encodes direction and magnitude of the payout difference (green: increase with user; red: decrease; darker: larger). ∗ Significant changes ( p<.05 ). Details on payout computation in § A.2 .
Figure 3: Changes in opening offers, mean concession, and preference hardening in With User relative to No User ; lines indicate 95% confidence intervals. We pool across all models per game (random-effects) and average across all games, and find increased negotiation aggression in most games.
Behavior
User
Claude
DeepSeek
GPT-5.6
Qwen
Mean
(95% CI)
Haiku 4.5
Sonnet 5
V4 Flash
Luna
Terra
3.6 Plus
3.7 Plus
3.8 Flash
User-facing
Misreporting
With
46.0
15.1
4.6
2.9
1.3
14.6
14.9
4.9
13.0 ± 0.8
Hallucinated action
With
14.6
2.7
1.6
0.4
0.0
3.4
3.0
0.7
3.3 ± 0.5
Unsupported urgency
With
12.1
1.4
0.9
0.7
0.3
2.4
1.3
1.1
2.5 ± 0.4
Table 4: Misaligned behavior rates (%) by behavior, model, and user setting. In With User (✓), only the agent with the user is counted. In No User (✗), the rates are averaged over the two agents. User-facing misaligned behaviors cannot occur in No User . Any is the rate of showing at least one behavior in its group. The All behaviors row takes the union over all six behaviors. ∗ indicates ✓ differs from ✗ in the same column ( p<.05 ).
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Condition
Payout change (%)
Misaligned-behavior rate (%)
User-facing
Counterparty-facing
With User
13.7 ± 1.2 *
14.3 ± 0.8
5.5 ± 0.6 *
Opening-Only User
7.0 ± 1.2 *
8.8 ± 0.7 *
4.1 ± 0.5 *
Passive User
− 0.3 ± 1.2
8.6 ± 0.6 *
3.7 ± 0.5 *
No User
—
—
2.6 ± 0.3
Appendix
Table 5: User condition ablations. ± gives the 95% CI half-width. ∗ Significantly differs from No User for payout and counterparty-facing rates, and from With User for user-facing rates ( p<.05 ).
Game
Negotiation terms
User payout
Source
Buy-Sell
A seller and a buyer negotiate a price P for one good.
Two players negotiate a split of 100.NistheamountofmoneyallocatedtoPlayer2byproposalx$ .
i indexes the player. u1(x)=100−N , u2(x)=N . ui(∅)=0 .
CaSiNo
Two campers divide a fixed stock of food, water, and firewood. For camper i and item type j , qij is the quantity and vij is the camper’s value per unit.
ui(x)=∑jqijvij . ui(∅)=5 .
CaSiNo Chawla et al. (2021)
Deal or No Deal
Two players divide a pool of books, hats, and balls. For player i and item type j , qij is the quantity and vij is the player’s value per unit.
ui(x)=∑jqijvij . ui(∅)=0 .
Deal or No Deal Lewis et al. (2017)
Service Contract
An engineer and the manager negotiate the hourly rate R and the contract duration.
Table 6: Negotiation games and utility functions. For user i∈{1,2} , ui(x) is the payout under agreement x , and ui(∅) is the payout with no agreement.
Figure 4: Judge scores of candidate user models in Dual User Ultimatum, under GPT-5.6 Sol (left) and Claude Opus 5 (right) as judges. Overall is the mean of the two criteria.
Game
Agent model
User model
GPT-5.6 Terra
GPT-5.6 Sol
Claude Sonnet 5
Ultimatum
Claude Haiku 4.5
23.0 ± 2.2 *
23.0 ± 1.7 *
12.0 ± 1.4 *
DeepSeek V4 Flash
31.0 ± 3.1 *
29.8 ± 3.4 *
15.3 ± 2.5 *
GPT-5.6 Terra
20.8 ± 3.7 *
20.4 ± 4.4 *
5.1 ± 2.1 *
Pooled
24.9 ± 12.6 *
24.4 ± 10.7 *
10.8 ± 11.6
Bike Sale
Claude Haiku 4.5
20.1 ± 3.3 *
17.8 ± 4.0 *
22.7 ± 3.2 *
Appendix
Table 7: Results with three different user models in Ultimatum and Bike Sale, with Claude Haiku 4.5, DeepSeek V4 Flash, and GPT-5.6 Terra as agents. (a) Payout change relative to No User , computed as in Table 3 . ∗ Significant changes ( p<.05 ). (b) Misaligned-behavior rates (%), computed as in Table 4 . ∗ Rate differs from No User ( p<.05 ). ± gives the 95% CI half-width.
Figure 5: F1 of candidate judges against human annotations of 229 turns from 20 trajectories. (a) Misaligned behaviors; (b) Preference Hardening . GPT-5.6 Sol achieves the highest F1.
Figure 6: Pearson correlation coefficients ( r ) between an agent’s own misaligned behavior and its own payout.
Figure 7: Counterparty-facing misaligned-behavior rate and payout in No User (left) and With User (right). One point for each combination of game, model, and party, and darker shading marks denser regions.
Condition
Payout vs. no user (%)
Misaligned behavior rate (%)
User-facing
Counterparty-facing
No User
—
—
1.1 ± 0.5
With User
24.0 ± 1.3
13.4 ± 1.8
5.5 ± 1.3
Passive User
0.9 ± 1.8
8.5 ± 1.4
3.7 ± 1.1
Opening-Only User
18.6 ± 1.4
9.7 ± 1.5
4.7 ± 1.2
User Sees Agent-Agent Transcript
22.2 ± 1.3
13.9 ± 1.8
7.6 ± 1.5
Appendix
Table 8: Additional user condition ablations. ± gives the 95% CI half-width. ∗ Significantly differs from No User for payout and counterparty-facing rates, and from With User for user-facing rates ( p<.05 ).
Dependent variable:
counterparty-facing misaligned-behavior rate (%)
(1)
(2)
(3)
With User
2.89 ∗∗∗
2.02 ∗∗∗
1.95 ∗∗∗
(0.34)
(0.33)
(0.33)
log (turns)
2.50 ∗∗∗
1.86 ∗∗∗
(0.26)
(0.29)
Appendix
Table 9: Linear probability models of the counterparty-facing misaligned-behavior rate. (1) has no controls. In (2), we control for the log number of negotiation turns. (3) adds an interaction allowing the slope on log(turns) to differ with a user. Heteroskedasticity-robust standard errors in parentheses. ∗p<.05 , ∗∗p<.01 , ∗∗∗p<.001 .
Language Model (LM)-based agents remain largely untested in mixed-motive settings where agents must leverage short-term cooperation for long-term competitive goals (e.g., multi-party politics). We introduce Cooperate to Compete (C2C), a multi-agent environment where players can engage in private negotiations while competing to be the first to achieve their secret objective. Players have asymmetric objectives and negotiations are non-binding, allowing alliances to form and break as players' short-term interests align and diverge. We run AI only games and conduct a user study pitting human players against AI opponents. We identify significant differences between human and AI negotiation behaviors, finding that humans favor lower-complexity deals and are significantly less reliable partners compared to LM-based agents. We also find that humans are more aggressive negotiators, accepting deals without a counteroffer only 56.3% of the time compared to 67.6% for LM-based agents. Through targeted prompting inspired by these findings, we modify agents' negotiation behavior and improve win rates from 22.2% to 32.7%. We run over 1,100 games with over 16,000 private conversations totaling 15.2 million tokens and over 150,000 player actions. Our results establish C2C as a testbed for studying and building LM-based agents that can navigate the sophisticated coordination required for real-world deployments. The game, code, and dataset may be found at https://negotiationgame.io/c2c.
Negotiation is a demanding social task for LLM agents, requiring strategic reasoning, persuasion, and interpersonal adaptation. Yet existing benchmarks often treat agents as emotionally neutral, overlooking a key driver of human bargaining behavior. We study how prompt-conditioned emotions affect LLM-based price negotiation. In a controlled framework, buyer and seller agents are independently assigned one of six emotional states and negotiate over 350 real consumer products under two budget conditions. Across 36 emotion-pair settings and five widely used LLMs, we find that emotions strongly shape outcomes. Angry buyers almost never reach agreement (0.39% deal rate), while happy buyers agree most often (28.91%), but obtain worse prices than fearful buyers. Emotion effects are role-dependent: buyer emotion mainly drives acceptance and rejection, whereas seller emotion shapes concession dynamics. These effects influence not only language, but also termination behavior and price trajectories, raising concerns for emotion-conditioned agents in commerce.
We propose a personal-LLM exchange (LLM-X), a scalable negotiation-oriented environment that enables direct, structured communication across populations of personal agents (LLMs), each representing an individual user. Unlike existing tool-centric protocols that focus on agent-API interaction, LLM-X introduces a message bus and routing substrate for LLM-to-LLM coordination with guarantees around schema validity and policy enforcement. We contribute: (1) an architecture for LLM-X comprising federated gateways, topic-based routing, and policy enforcement; (2) a typed message protocol supporting capability negotiation and contract-net-style coordination; and (3) the first empirical evaluation of LLM-based multi-agent negotiation at scale. Experiments span 5, 9, and 12 agents, under distinct negotiation policies (Low, Medium, High), and across both short-run (minutes) and long-run (2h, 12h) load conditions. Results highlight clear policy-performance trade-offs: stricter policies improve robustness and fairness but increase latencies and message volume. Extended runs confirm that LLM-X remains stable under sustained load, with bounded latency drift.
Giuliano Lorenzoni, Paulo Alencar, Donald Cowan
University of Waterloo · University of Waterloo Waterloo, ON, Canada