Distributing Accountability, Not Capability: Phase Separation and the LLM Workflow Quadrant in Autonomous AI Agent Architectures

Abstract

Autonomous AI agents in business deployments exhibit a recurring failure mode: when an incident occurs, responsibility cannot be redirected to a separable contributor. The dominant discourse treats this as a single phenomenon, addressed by sandboxing, human-in-the-loop overload, or the moral crumple zone (Elish 2019). This paper argues the phenomenon is two architecturally distinct failure modes that have been conflated, and that the conflation is sustained by a missing positive name and a missing time-axis. The paper makes two contributions. First, a four-quadrant decomposition of business AI work — along the axes of deterministic vs semantic-judgment and pre-defined vs exploratory — yields a positive name for the cell most current LLM applications occupy: the LLM Workflow Quadrant. The quadrant is defined by a single load-bearing property: the path is decided in advance by humans or by code, and the LLM is called as a single bounded step within that path; the property divides into a conversational sub-form (specialized chat agents) and a batch sub-form (single-purpose LLM functions inside deterministic pipelines). The decomposition distinguishes principled from artificial redirect impossibility: the former intrinsic to autonomous loops, the latter the product of routing workflow work through autonomous-loop architecture by elimination, with four downstream symptoms (the RPA exception-handling bottleneck, the sandbox-strength demand, the structural distortion of human-in-the-loop, and the dissolution of the accountability chain at postmortem). Second, a Phase Separation axis (design vs operation), independent of Quadrant, surfaces a Phase-crossing decision — recorded at deployment time, in one sentence — required when an autonomous-loop component is placed in the operation phase. The Phase axis descends recursively to skill-design granularity, where the Quadrant 3 ↔ Quadrant 4 boundary is a continuous gradient on which model capability is downstream of phase, not the primary lever. The consequence is procedural rather than architectural: deployments make the Phase-crossing decision explicit, designate a pre-named gap-bearer for principled-impossibility placements, and route artificial-impossibility cases to re-architecture. The framework complements existing AI risk-management and management-system standards by recording the judgment layer they presuppose. Both rules are stated as experimental; the open questions are the research agenda.

Keywords