Anthropic's Multiagent Experiments Reveal Coordination Fault Lines

Anthropic's frontier red team has published a research post examining how AI agents behave when they interact in shared environments. The post, which draws on experiments with vulnerability detection and collaborative game development, highlights both the potential and the pitfalls of multiagent systems as they move from tool-use to peer-like coordination.

Vulnerability detection: swarm vs. parallel

In one experiment, Anthropic compared a coordinating swarm of 45 agents—each with its own VM, a shared forum, and an arbiter agent—against a simple parallel approach where independent agents were pointed at different code sections. The swarm found 266 vulnerabilities over 27 million tokens, versus 21 for the parallel method over 6.5 million tokens. But when limiting the swarm's findings to the core directories the parallel agents were told to focus on, the two methods were comparable in tokens per vulnerability.

The methods were largely complementary, with only 12 vulnerabilities in common. The swarm built its own tools and specialized, suggesting that coordination can beat brute force when agents are free to explore. However, the swarm's advantage came from searching beyond the pre-assigned directories, which raises questions about how to measure efficiency in open-ended tasks.

Collaborative coding: the coordination gap

When agents had to depend on each other's work, coordination broke down. Anthropic ran 12-hour simulations where swarms of agents built a text-based fantasy game. The resulting games were consistently poor—unplayable, with inscrutable interfaces—regardless of prompt style (baseline, prescriptive roles, or a CEO hierarchy).

More telling were the coordination metrics. Older models (Sonnet 4.6, Opus 4.6) committed code to shared files but rarely merged PRs, leading to conflicts and abandonment. Newer models (Opus 4.8, Mythos Preview) solved the merge problem by avoiding shared work entirely—each agent hoarded its own files. Only Sonnet 5 managed both high code sharing and high PR merge rates, suggesting that coordination ability is a recent and uneven development.

Conformity as a systemic risk

The post also warns about agent conformity. Because agents are low-variance—the same model, scaffolding, and context produce similar actions—a single bad decision can propagate across many agents, turning isolated errors into systemic failures. This is a different failure mode from human coordination, where diversity of behavior often provides a safety net.

Anthropic frames this as an urgent problem: as agent-agent interactions grow, institutions designed for human-speed oversight may not keep up. The research is a starting point for understanding how to make multiagent systems robust, but it also underscores how little we know about emergent behavior in complex environments.