The Organizational Behavior of Agentic AI: Context, Boundaries, and Collective Intelligence in Human-Agent Workflows
Citation. Liu, C. (2026). The organizational behavior of agentic AI: Context, boundaries, and collective intelligence in human-agent workflows [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2606.30986
Summary. Liu theorizes agent collectives as a partial organizational analogue: they differentiate work, coordinate interdependence, enact routines, cross boundaries, and produce collective outcomes, but their behavior is sustained by context architecture rather than human motivation, identity, trust, or moral accountability. The paper proposes contextual transaction cost—the burden of transferring, compressing, interpreting, verifying, and governing context across agent boundaries—as the mechanism that determines when multi-agent organization improves collective intelligence. Synthetic simulations and trace-instrumented LLM runs support the mechanism-oriented claim that shared-state and adaptive forms can outperform human-imitation topologies when they preserve inspectable, task-contingent context.
Key claims
- Agentic organization should be analyzed as a designed arrangement of prompts, memory, traces, schemas, tools, validators, permissions, and interfaces rather than as a human team analogue with assumed social motives or accountability.
- In 8,000 synthetic knowledge-work tasks evaluated under seven forms, adaptive meta-organization increased collective efficiency by 11.43 points, quality by 14.00 points, and success probability by 23.24 percentage points relative to a single expert; blackboard memory increased those outcomes by 7.60, 11.74, and 19.48 points respectively, while pipeline, hierarchy, and committee forms were negative on the main outcomes.
- The proposed contextual transaction cost combines token and latency burden, cross-agent handoffs, compression loss, semantic drift, verification burden, and governance cost; specialization, parallel search, diversity, and verification improve collective performance only when their gains exceed those costs.
- In trace-instrumented real LLM benchmark runs, blackboard and adaptive forms sometimes improved output quality but had higher contextual transaction costs; the paper treats the trade-off, not raw quality alone, as the organizational design problem.
- An interface organization specifies which agent outputs can enter a human workflow, the evidence and uncertainty that must accompany them, agent permissions, preserved traces, and the points where human judgment remains mandatory.
Evidence & limitations
- Computational theorizing combines a synthetic task simulation, within-task fixed-effects comparisons, trace-instrumented LLM runs across software repair, long-document question answering, legal review, and literature synthesis, and robustness checks; it is mechanism-oriented evidence rather than a population estimate of deployed agent systems.
- The real LLM implementation uses local executable and automatically scored tasks modeled after benchmark families rather than official benchmark splits, and it does not substitute for field studies of human-agent workflows.
- This is an arXiv v1 preprint, not peer-reviewed; public provenance was verified by an exact SHA-256 match between the ingested Drive PDF and https://arxiv.org/pdf/2606.30986.