Why negative cases are the key to knowing when coordination works
Negative cases—situations where coordination fails, underperforms, or adds overhead—are not just failures to hide; they are the data that tell you when coordination is worth the complexity. A controlled study across 260 configurations found that the single most robust predictor of whether multi-agent coordination helps or hurts is the baseline performance of a strong single agent [1]. When that baseline is already high, adding agents is unlikely to improve results—and can even amplify errors, a failure mode the authors call 'baseline-scaled error amplification' [1]. This means that without reporting negative cases, you cannot build a reliable rule for when to use coordination at all.
The same lesson appears from a different angle in a benchmark for open-ended coordination: current LLM agents average only about 6% normalized return, but their failures are not uniform [3]. Some agents achieve strong base-task reward yet much lower coordination reward, showing that individual competence does not translate to coordination competence [3]. If papers only reported the successes, readers would never learn that coordination is a distinct bottleneck—one that needs its own evaluation and design.
What negative cases reveal about coordination mechanisms
Negative cases also expose which coordination mechanisms actually matter. In the open-ended benchmark, ablations showed that communication is the largest contributor to coordination success, while memory and reasoning help only when used to maintain multi-step plans [3]. This suggests that when coordination fails, the cause is often a lack of communication, not a lack of raw intelligence. Similarly, a study on device-cloud collaboration found that a step-level coordinator could achieve 92.5% of the cloud-only success rate while using cloud in only 46.3% of steps—but this success depended on careful routing decisions [4]. Without reporting cases where routing fails, you cannot know when the trade-off between performance and cost is acceptable.
Another paper, SyncPlan, explicitly addresses the trade-off between efficiency and adaptivity: one-shot planning reduces overhead but produces plans that can quickly become stale, while repeated LLM invocations introduce latency [5]. Their solution uses explicit synchronization and adaptive correction, achieving state-of-the-art success rates with less than 0.05% of the wall-clock runtime of existing coordinators [5]. The negative case here is the failure of open-loop plans—without reporting that, the need for synchronization would be invisible.
How to report negative cases so they're actually useful
To be useful, negative cases must be reported with the same rigor as positive ones: specify the exact conditions, the baseline, and the failure mode. The controlled study in [1] provides a model: they held task prompts, tools, and compute budgets constant while varying only coordination structure and model capability, and they derived a predictive model that selects the best architecture in 87% of held-out configurations [1]. This kind of reporting turns a negative result into a generalizable rule.
Similarly, the SAGA paper on long-horizon strategy games reports that current LLM agents fall short in three specific ways—inferring spatial relations, allocating resources, and improving from delayed feedback—and then designs mechanisms to address each [2]. By naming the failures, they make the improvements interpretable. In contrast, a benchmark on long-horizon attacks found that defenses designed for single-turn interactions fail to reliably mitigate long-horizon threats [6]. Reporting that negative result is essential for the community to know that existing defenses are insufficient.
Finally, negative cases should be reported even when they are uncomfortable. The LLM-Coordination study found that LLM agents can outperform RL baselines in some coordination tasks, but also that they struggle with explicit assistance—proactively helping partners at the cost of their own task progress [7]. That negative finding is as valuable as the positive one, because it points to a specific capability gap that future work must address.
About These Sources
This answer is built on 7 studies (1 peer-reviewed, 6 preprints) — published from 2023 to 2026, 6 from 2024 or later, 1 in Q1 journals — selected as the most relevant from 7 studies that passed quality screening, drawn from 62 papers retrieved from a database of over 500 million.
Sources used in this answer
Capable language models can outgrow the benefits of collaboration
In a controlled study across 260 configurations, single-agent baseline performance was the most robust predictor of whether multi-agent coordination improves or hurts, with a capability-saturation threshold beyond which adding agents is unlikely to help; the model selected the best architecture in 87% of held-out configurations.
SAGA: Scene-Aware, Goal-Evolving Agents for Long-Horizon Strategy Game Planning
SAGA, an LLM multi-agent framework for long-horizon strategy games, addresses three specific weaknesses (spatial inference, resource allocation, and delayed feedback) and outperforms five LLM baselines on mean final score in CivRealm, with 27% fewer output tokens.
Benchmarking Open-Ended Multi-Agent Coordination in Language Agents
In a benchmark for open-ended multi-agent coordination, 13 modern LLMs averaged only ~6% normalized return, with failures not uniform; ablations showed communication is the largest contributor to coordination success, while individual task competence does not imply coordination competence.
Hera: Learning Long-Horizon Coordination for Device-Cloud Collaborative LLM Agents
Hera, a step-level device-cloud coordinator, achieved 92.5% of the cloud-only success rate while using cloud in only 46.3% of steps across ALFWorld, WebShop, and AppWorld, using imitation learning followed by cost-aware reinforcement learning.
SyncPlan: Long-Horizon LLM Coordination with Explicit Synchronization and Adaptive Correction
SyncPlan, a plan-execute-correct framework with explicit synchronization and adaptive correction, achieved state-of-the-art success rates on Overcooked and Honor of Kings while using less than 0.05% of the wall-clock runtime of existing LLM-based coordinators.
AgentLAB: Benchmarking LLM Agents against Long-Horizon Attacks
AgentLAB, the first benchmark for long-horizon attacks on LLM agents, found that representative agents remain highly susceptible to five attack types across 28 environments and 644 test cases, and that single-turn defenses fail to reliably mitigate long-horizon threats.
LLM-Coordination: Developing Coordinating Agents with Large Language Models
The LLM-Coordination framework showed that LLM agents can outperform RL baselines in sustained coordination and robustness to partners, but struggled with explicit assistance—proactively helping partners at the cost of their own task progress—in Overcooked-AI.