A useful new paper challenges a common assumption: adding more agents does not automatically improve outcomes.
In The Organizational Behavior of Agentic AI (July 2026), Canhui Liu describes agentic AI as a “partial organisational analogue”. Agent collectives can divide labour, coordinate tasks, conduct reviews, maintain memory, and produce collective outcomes. But unlike human organisations, they are not driven by trust, identity, incentives, authority, professional norms, or moral responsibility. Their behaviour is shaped by prompts, memory structures, tools, permissions, validation mechanisms, and context architecture.
The most practical contribution is the concept of contextual transaction cost (CTC). When AI agents exchange work, there is a cost attached to transferring, compressing, validating, and interpreting context. Excessive hand-offs can introduce semantic drift, verification overhead, and error propagation. According to the paper, these costs often determine whether multi-agent systems outperform a single model.
The findings challenge a common assumption in AI governance and solution design: adding more agents does not automatically improve outcomes.
Simulated and trace-based evaluations found that some human-inspired structures such as committees, hierarchies, and sequential pipelines frequently underperformed when they created unnecessary context transfers. By contrast, shared-memory and adaptive orchestration models often performed better because they reduced context loss and made evidence more durable and auditable.