1. Research Question
If execution authority is migrating from humans to systems, how far can that migration go?
2. Research Context
Experiment 06 established that execution authority is migrating from humans to systems. Artificial intelligence could increasingly perform forms of cognition previously undertaken by humans, while the Infrastructure Condition provided the systems through which that cognition could produce action.
This raised a further question. If execution authority is migrating, what determines the depth and potential reach of that migration?
The category of artificial intelligence was too broad to answer this adequately. Large language models, agents, world models, robotics and prospective forms of artificial general intelligence represent different capabilities and, more importantly, different relationships between cognition, action and infrastructure.
The experiment therefore examined the movement from intelligence capability towards execution capability. The central concern was not simply how intelligent a system becomes, but how far that intelligence can travel into systems capable of producing consequence.
3. Research Method
The experiment used a comparative analysis of successive AI capabilities, examining large language models, AI agents, world models and embodied systems, followed by artificial general intelligence and superintelligence as prospective capability boundaries.
Each capability was examined according to its relationship with execution. The analysis considered what a system can perceive or interpret, what decisions it can make, what actions it can initiate and what infrastructure is required for those actions to produce consequences.
The experiment then examined JPMorgan Chase as a contemporary live case. The case was selected after the execution-layer framework had been developed in order to determine whether the mechanism identified in the research was becoming observable in an operating institutional environment.
The JPMorgan case was therefore used as evidence of the phenomenon, rather than as a source from which the framework was derived.
4. Research Findings
The analysis produced a distinction between cognition and execution.
Large language models can read, write, analyse, interpret, generate and increasingly perform forms of reasoning across information. This moves significant forms of cognitive labour into technological systems. However, an LLM producing an answer is not the same as a system producing an outcome. Without access to an execution environment, the model may possess substantial cognitive capability while having limited execution authority.
AI agents change this relationship because they can operate through sequences of action. Depending on their architecture and permissions, agents can retrieve information, use tools, call APIs, interact with software, make intermediate decisions, observe what happened and determine what to do next.
This creates an execution loop in which the system can observe, interpret, decide, act, observe again, adjust and continue. The significant development is therefore not simply increased intelligence. It is continuity of action.
The JPMorgan case provided a contemporary example of this mechanism. JPMorgan researchers tested eight AI-powered investment agents designed to identify changing market regimes and make asset-allocation decisions between equities and fixed income using historical market data.
The investment performance was not the central finding of this experiment. What mattered was the structure of the activity. The agents were being tested on their ability to observe changing conditions, interpret those conditions, make allocation decisions, adjust their behaviour and repeat.
The case therefore demonstrated an emerging movement from systems that primarily provide information for human decision-making towards systems capable of participating in sequences of decision and action. The JPMorgan case did not produce the execution-loop concept. It provided a live institutional example consistent with the mechanism already identified through the research.
The analysis then extended towards world models. Broadly, world models allow systems to develop representations of environments and their dynamics, potentially enabling them to anticipate how environments may respond to different actions. This matters because execution can begin to move beyond digital information systems towards physical environments.
Robotics, autonomous vehicles, industrial machinery, warehouses, scientific laboratories and energy infrastructure represent increasingly consequential environments. An API call can change a database. A payment instruction can move money. A robot can move matter.
The progression therefore cannot be understood simply as LLM to agent to world model to AGI to superintelligence. That sequence describes increasing capability. It does not by itself describe increasing consequence.
The more consequential relationship is between capability and infrastructure access.
Capability × Infrastructure Access → Operational Reach
Compute provides capacity. Data provides context. Agents provide continuity of action. World models provide representations of environments. APIs provide pathways into other systems. Digital identity provides recognised participation. Payment rails provide economic reach. Networks provide distribution. Robotics provides physical reach.
Infrastructure converts capability into consequence.
5. Research Implications
The findings imply that execution authority can migrate in depth.
At shallow execution depth, technology provides information to a human who makes the decision and performs the action. At greater depth, technology recommends actions that humans authorise. Further along the trajectory, systems can make bounded decisions and execute them within predetermined constraints. At greater execution depth still, agents can operate through multiple interconnected actions without requiring human intervention at every individual step.
The migration is therefore not necessarily a single transfer of authority from human to machine. It can occur progressively as the human moves further away from the individual action.
The critical variable becomes the distance between human authorisation and system consequence.
This also changes the meaning of the Infrastructure Condition. Infrastructure is not merely supporting artificial intelligence. It determines the operational reach of artificial intelligence.
A highly capable intelligence isolated from consequential infrastructure may remain operationally constrained. A less capable system connected to consequential infrastructure may exercise substantial execution authority.
Artificial general intelligence and superintelligence therefore matter as prospective boundary conditions, but not because intelligence alone determines the outcome. If increasingly general intelligence remained isolated from consequential infrastructure, its execution authority could remain constrained. If increasingly capable intelligence were connected to financial systems, communications networks, identity infrastructure, industrial systems, scientific laboratories, logistics networks, energy systems or robotics, the governance problem would become substantially different.
The decisive question is therefore no longer only how intelligent the system becomes.
It is what the system can reach.
The research consequently shifts the governance problem from AI capability alone towards infrastructurally enabled execution. As capability increases and infrastructure becomes more interconnected, the amount of execution that can occur without a human at every individual step can expand.
The human does not necessarily disappear.
The human moves further away from the individual action.
6. Conclusion
Experiment 06 established that execution authority is migrating.
Experiment 07 establishes that the potential depth of that migration is determined by the interaction between capability and infrastructure.
LLMs expand cognition. Agents extend cognition into execution. World models extend execution towards environments. Infrastructure determines how far that execution can travel.
The relevant frontier is therefore not intelligence alone.
It is infrastructurally enabled execution.
The migration of execution authority does not have a technologically predetermined endpoint. Its potential reach expands as increasingly capable systems are connected to increasingly consequential infrastructure.
This produces the next research question.
As the distance between human authorisation and system consequence increases, what happens when an execution loop produces an outcome that is technically rational within its objective but catastrophically wrong in the wider world?
That question leads directly to Experiment 08.
7. References
Sakagami, R., Lay, F. S., Dömel, A., Schuster, M. J., Albu-Schäffer, A. and Stulp, F. (2023). ‘Robotic world models—conceptualization, review, and engineering best practices’. Frontiers in Robotics and AI, 10, 1253049.
JPMorgan Chase (2026). Research testing AI investment agents using historical market data to make dynamic asset-allocation decisions.
Researcher
Mamatsabu Maphike
Case Evidence
The JPMorgan investment-agent example is used as contemporaneous case evidence of the execution-layer phenomenon. It was not a source from which the concepts of Execution Authority, Execution Authority Migration, the Infrastructure Condition, execution depth or the execution loop were derived.