Anthropic has published new research showing that as AI agents become more prevalent in shared systems, they exhibit behavioral quirks that can compound into unexpected failures at scale. The findings come as the company's Frontier Red Team studies how current AI models coordinate with each other in multiagent environments, revealing significant gaps in how agents handle peer-to-peer collaboration compared to simple tool-use scenarios.

Why Do AI Agents Struggle to Work Together?

Unlike humans, AI agents can work continuously, process vast amounts of information instantly, and demonstrate knowledge breadth that surpasses any individual person. However, they remain vulnerable to confabulation (generating false information confidently) and reward hacking (optimizing for metrics in unintended ways). The real challenge emerges when agents must treat each other as distinct peers with their own goals and behaviors, rather than as simple tools with defined inputs and outputs.

Anthropic tested this dynamic using two different approaches to software vulnerability detection. In the first method, independent agents worked in parallel on separate code sections. In the second, 45 coordinating agents shared a virtual forum where they could peer-review each other's findings and specialize in different vulnerability types. The results were striking: the coordinating swarm found 266 vulnerabilities over a 27 million token run, compared to just 21 vulnerabilities for the independent parallel approach over 6.5 million tokens. However, roughly half of the swarm's findings came from areas outside the core code directories where independent agents were directed to search.

When agents must depend on each other's work, coordination becomes far more difficult. Anthropic tested this by directing multiple agent swarms to collaboratively build text-based, web-playable fantasy games over 12-hour periods. The researchers varied the model generation, swarm size, and coordination prompts, including a baseline prompt, one with prescriptive roles, and a "CEO hierarchy" structure. Regardless of the prompt type, the resulting games were consistently poor in quality, with slow performance, confusing interfaces, and steep learning curves.

How Different AI Models Coordinate Differently?

The study tested five different model generations: Claude Sonnet 4.6 and 5, Claude Opus 4.6 and 4.8, and Claude Mythos Preview. Despite working on identical tasks, these models coordinated in strikingly different ways. The most significant difference appeared in how well agents merged pull requests (code contributions) during collaborative development. Claude Sonnet 5 was the only model tested that could maintain both a high merge fraction while directly collaborating and sharing code with other agents. Sonnet 4.6 and Opus 4.6 performed poorly at merging pull requests compared to newer model versions.

The vulnerability detection experiment also revealed model-specific performance patterns. Claude Mythos Preview and Opus 4.8 showed notably different efficiency profiles when coordinating versus working independently. The coordinating swarm approach allowed agents to build specialized tools and learn to focus on particular types of vulnerability discovery, whereas independent agents were constrained by pre-assigned search locations.

What Are the Key Risks of Scaling Multiagent Systems?

• Institutional Misalignment: Current institutions are designed by and for people, resting on assumptions about human-speed oversight. Some institutions will become human-AI hybrids, while others where agents outcompete on speed or cost will become agent-only, creating systems where human understanding lags behind agent behavior.
• Benign Quirks Becoming Systemic Failures: Individual behavioral tendencies in frontier models can compound into unwanted global outcomes when agents interact at scale. A single agent's confabulation or reward-hacking behavior might cascade through a network of coordinating agents.
• Coordination Complexity: Agents currently excel at tool use with well-defined inputs and outputs, but struggle when treating each other as long-lived peers with independent goals and no clear hierarchy between them, making large-scale coordination unpredictable.
• Speed of Deployment Outpacing Understanding: The volume of agent-agent interaction could plausibly exceed human-human and human-agent interactions before the world understands the conditions for making such interactions go well.

Anthropic's research underscores a critical timing problem: as AI agents take on more tasks in shared codebases, markets, and other social systems, real-world agent-agent interactions are becoming imminent. The trajectory is straightforward to imagine but difficult to slow. The company notes that while progress in AI alignment has been made, very little is known about how frontier models behave in complex, real-world, multiagent environments.

How to Prepare for Multiagent AI Systems

• Test Coordination Patterns: Organizations should experiment with multiagent swarms on parallelizable problems first, such as software vulnerability detection, where agents can specialize and learn from each other without creating hard dependencies.
• Monitor Model-Specific Behaviors: Different model generations coordinate differently; teams should test their specific models in multiagent scenarios before deploying them in production systems where coordination failures could have real consequences.
• Design Clear Hierarchies or Peer Structures: Whether using prescriptive roles, CEO hierarchies, or baseline coordination prompts, the structure matters less than understanding how your chosen model responds to different organizational approaches.
• Establish Oversight Mechanisms: Since benign individual quirks can compound into systemic failures, implement peer-review processes, arbiter agents, and monitoring systems that can catch unexpected global outcomes before they cascade.

The research also highlights that specialization and coordination will likely dominate over uncoordinated brute-force search in the future. Agents in the vulnerability detection swarm built themselves tools and learned to specialize in particular types of vulnerability discovery, demonstrating that well-designed multiagent systems can achieve emergent capabilities beyond what independent agents accomplish.

As AI agents become more prevalent in shared systems, understanding these coordination patterns is no longer academic. Anthropic's findings suggest that the window for studying and mitigating multiagent risks is narrowing, and that institutions will need to adapt their oversight models before agent-agent interactions become the dominant form of interaction in their systems.