Abstract
Machine Aliveness studies machines that act with initiative, remain grounded in a body, retain history, and live alongside people. This paper examines one part of that agenda: how such machines can continue to develop through experience in physical and social worlds that do not reset around them. A machine that lives in the real world does not encounter a sequence of independent tasks. It accumulates history, changes the conditions of later interaction, and returns to a world already altered by what happened before.
Reality is therefore indispensable, but insufficient as the only substrate of development. It provides committed consequences and external evidence, yet it cannot be freely forked, replayed, accelerated, or restored. A household cannot be duplicated from the same instant; a person cannot be reset to a prior experience; months of interaction cannot be rerun under controlled counterfactuals.
We call the missing computational layer the Second Loop. The First Loop unfolds in reality. The Second Loop makes the conditions of future development computational: lived experience can inform world reconstructions; candidate interactions can be branched and compared; models of people, processes, and environments can be revised; and the machinery that constructs later developmental situations can itself be retained and improved. The Second Loop does not replace reality. It expands what can be learned from it while preserving the distinction between internally generated evidence and external experience.
We propose Agents as World (AAW) as an organization for this computational world. The underlying object is a World Process: a persistent computational process through which participants, events, physical consequences, and histories continue to unfold. Actor Agents participate from local perspectives; World Agents construct, coordinate, and revise ongoing processes; and a tool-grounded World Kernel acts as a commitment layer, separating proposed futures from committed history.
For human-centered physical intelligence, people are not merely sources of demonstrations, instructions, or feedback. They are persistent participants whose memory, intentions, routines, and relationships evolve through interaction and become part of the world a machine encounters again.
This is one route from machine intelligence toward machine aliveness: not simply more capability, but continued development across a lived history, supported by computational worlds that can remember, branch, change, and remain grounded in reality.
The position
The First Loop unfolds within a world. The Second Loop revises the processes through which later worlds unfold.
1. Development beyond the episode
1.1 The world continues
Embodied intelligence is often organized around episodes: a task begins, an agent acts, a trajectory ends, and learning follows. But a machine living in the real world does not encounter such discrete episodes.
The world continues. Objects remain moved. People remember. Routines shift. Expectations form. The machine retains new memories, models, tools, and behaviors. What happens now shapes what follows.
We use development for this broader process of persistent change. Learning is part of it, but development also includes changes in memory, interaction strategy, models of people and environments, and the machinery through which future experience is produced.
1.2 The First Loop is reality
For physical intelligence, reality is the First Loop: the place where actions acquire real consequences and internal assumptions meet external evidence.
But reality does not branch easily. A household cannot be forked from the same moment; a person cannot be restored to an earlier experience; a month of interaction cannot be replayed under alternative conditions. Physical time, safety, hardware, and human participation make counterfactual exploration costly.
Reality is the most authoritative source of evidence, but a costly place to explore developmental paths.
1.3 The Second Loop
We call the missing computational layer the Second Loop. The First Loop is where consequences become real. The Second Loop is where the conditions of future development become computational.
Within it, lived experience can be reconstructed, alternative futures can be explored, models of people and environments can be revised, and the processes that generate later developments can themselves change.
The Second Loop does not replace reality. It expands what can be learned from it.
2. The world as a continuing process
2.1 The world side must be executable and revisable
Once the world becomes part of machine development, a passive environment or observation predictor no longer suffices. We call the object of interest a World Process: a persistent computational process through which participants, events, physical consequences, and histories continue to unfold over time.
A World Process is an instantiated, running world—not merely a scene or a model of one. The Second Loop is the broader computational regime in which such processes can be instantiated, continued, branched, and compared, while the mechanisms that produce later World Processes can themselves be revised.
A World Process specifies more than what exists. It determines what continues unobserved, what each participant can perceive, how physical and human dynamics interact, how time advances, and which mechanisms govern its evolution. Its implementation may combine neural models, programs, scene descriptions, process models, actor specifications, physics settings, state stores, and executable tools.
It need not be end-to-end differentiable. What matters is that the world can run, that its consequences can be inspected, and that the machinery governing later worlds can be revised.
2.2 The human is inside the loop
The human boundary
People are not merely interfaces to the world. They are persistent, adaptive parts of the world.
Generalist embodied systems already use humans extensively—as sources of human video and demonstrations, and as instruction-givers or interactive operators at execution time. [13] [14] These are essential forms of human grounding. In these roles, however, the person largely appears at the boundary of the learning system: supplying data, commands, corrections, or evaluation.
For machines that develop through repeated interaction with people, that boundary is too narrow. People are part of the dynamics the robot acts within and those that persist after an interaction. A household member does not reset after a task. They remember, form expectations, carry private information, change routines, become more or less receptive, and bring these changes into later encounters.
This changes the modeling role of the human. A demonstrator provides data; an instructor provides a command; an evaluator provides a judgment. A Human Agent inside the world instead carries state forward and participates in the causal history through which later interaction unfolds. Persistent social agents and shared human–robot worlds have precedents in Generative Agents and Virtual Community. [5] [6]
Our concern is not simply whether a model can predict what a person will do next, but how persistent human state becomes part of the developmental world a machine repeatedly encounters.
2.3 The world remembers and commits
Continuity matters only when past consequences remain causally available to later states. The learner therefore does not simply act in a world; its actions partly shape the world it will encounter again.
This persistence is causal, not merely archival. A physical action can alter what another participant observes; an interruption can shift attention; a social response can change the learner's next action. Physical, human, and interaction dynamics therefore accumulate within one history.
A continuing World Process must preserve enough of that history for its consequences to remain available to the future, while distinguishing possible futures from realized ones. Possible futures may branch or be revised; realized consequences become part of committed history.
The Second Loop may change how later worlds unfold. It must preserve the history from which they unfold.
2.4 World-making must expose its own machinery
Finally, the world-making process must expose the machinery through which worlds are produced. Changing only finished scenes or trajectories provides little leverage over the conditions of future development. Reusable construction procedures, process models, observation interfaces, event rules, human-dynamics models, and temporal policies must themselves be addressable objects.
The world side of the Second Loop must therefore expose not only states and outcomes, but the processes through which later worlds are constructed, continued, and revised.
3. Agents as World (AAW)
Agents do not need to simulate everything. They need to organize the systems that make a world run.
Agents as World (AAW) is our proposed organization for World Processes within the Second Loop. Agents do not replace physics, state, or simulation tools; they organize how World Processes are constructed, operated, inspected, and revised, while specialized substrates determine executable consequences.
Why agents? Part of the world is already agentic: people carry beliefs, memories, intentions, routines, and relationships that affect what happens next. Meanwhile, computational agents increasingly operate world-making tools—authoring environments, invoking simulators and renderers, modifying programs, inspecting state, coordinating other agents, and acting through embodied controllers. [15] [16] [17] [18]
The important shift is therefore organizational. Physics engines, renderers, databases, event systems, process models, and robot controllers remain specialized substrates; agents increasingly provide a common layer through which they can be composed and operated.
AAW assigns agents two roles. Actor Agents participate within a World Process from local perspectives. World Agents construct, operate, and revise the processes through which it unfolds. A shared World Kernel preserves authoritative state, executable consequences, and committed history.
The lifecycle is process-oriented: instantiate, unfold, revise, inherit.
3.1 Actor Agents: local perspectives that persist
An Actor Agent represents a participant whose behavior depends on a local perspective rather than omniscient world state. It may represent a person, robot, animal, or another decision-making participant.
Its defining properties are persistence and locality. Identity, goals, beliefs, attention, memories, capabilities, and relationships can carry forward, allowing earlier interactions to alter later conditions. Its actions may include movement, waiting, communication, cooperation, interruption, or changes of plan, but remain bounded by what the actor can observe and physically execute.
For human-centered physical intelligence, one subclass deserves explicit treatment: the human.
3.2 Human Agents: persistent human dynamics
A Human Agent models a persistent participant whose state both shapes and is shaped by interaction. Relatively stable characteristics may belong to its specification; goals, beliefs, attention, memories, relationships, routines, preferences, and private observations evolve at runtime.
The primary target is not isolated behavior but interaction dynamics: how one encounter changes the state from which later encounters begin. Human Agents must therefore remain physically situated and epistemically local; what they perceive depends on position, attention, available signals, and shared history, while physical actions pass through the same execution substrate as other actions.
They must also remain non-oracular. A simulated preference, correction, demonstration, or emotional response is a model-generated hypothesis, not external evidence about a real person's private state. Human Agents are models to be calibrated against reality, not participants to be optimized into compliance.
3.3 World Agents: construct, operate, revise
World Agents make the non-actor world programmable: construction, operation, and revision.
During construction, they interpret world specifications, compose assets and layouts, set physical properties, instantiate processes, and assemble executable tools. They may use scene generators, physics engines, Blender, code, procedural pipelines, or validation utilities; no single simulator backend defines AAW.
During operation, they maintain process state, schedule events, manage dependencies, respond to actor actions, and coordinate tools. The world continues beyond loaded geometry: appliances run, visitors arrive, objects change, and processes advance outside the learner's immediate focus.
During revision, new evidence or discovered discrepancies can alter scene specifications, process models, physical parameters, event rules, or the procedures through which later World Processes are constructed. Such revisions may remain local or become reusable across future worlds.
This division enables scale: agent reasoning need not occur at every physics step. Agents author and supervise processes that ordinary programs execute until a decision, interruption, or revision is needed.
3.4 The World Kernel: revision without rewriting history
Agents propose. Tools execute. The world commits.
World-making mechanisms must be revisable without making already-realized history negotiable. The World Kernel is therefore the authoritative layer for public state, observation boundaries, event ordering, tool execution, snapshots, branching, and replay.
Agents may propose that a person moves, a door opens, or a process begins; the corresponding simulator, program, or declared abstraction determines the resulting transition. Later mechanisms may change, but consequences already committed to the current World Process remain part of its history.
The Kernel also serves as an epistemic boundary. Actor and World Agents do not directly declare reality: simultaneous actions require arbitration, observations respect local access, snapshots preserve active processes and backend state, and reproducible branches pin the relevant mechanisms and randomness.
Future mechanisms may be revised. Committed history may not be rewritten.
3.5 Persistent time, selective detail
Long-horizon worlds must preserve causal continuity without paying fine-grained computational cost at every moment. AAW therefore separates world time from the resolution of computation used to represent it.
The World Process follows one causal timeline. A three-second action occupies three seconds regardless of the model calls or physics steps required to compute it; conversely, a quiet hour need not receive the same resolution as a dense human–robot interaction.
Coarse process models advance routine periods while preserving active conditions, memory, and opportunities for interruption. When interaction becomes consequential, the relevant interval expands into fine execution—speech, motion, attention, cooperation, interruption, and physical response. Resolution can change dynamically as unexpected events demand detail, after which realized outcomes update what follows.
Refinement can resolve unrealized futures, but cannot alter history already observed by participants.
The objective is long-horizon continuity without uniform computational detail.
3.6 Inheritance: retain how worlds are made
Memory retains what happened.
Inheritance retains how future worlds can be made.
Continual development should retain more than finished scenes and trajectories. It should retain what has been learned about how future worlds are made: actor models, process templates, temporal policies, evaluators, construction programs, and other reusable world-making mechanisms.
Inheritance operates across levels. An actor changing its state is part of a World Process; revising the model that produces such changes alters that process; revising how such models are constructed, tested, or selected alters the infrastructure from which later World Processes arise.
These changes must retain their epistemic meaning. A more compliant simulated human is not necessarily a better human model, and a more frequently sampled scenario is not therefore more common in reality. [10]
Development becomes cumulative when what is learned in one World Process changes the worlds available to the next. Inheritance is how the Second Loop carries that leverage forward.
4. Closing the loops: development across worlds
Coupling computational worlds to lived reality
The First and Second Loops are two coupled regimes of machine development.
Reality provides what a computational world cannot manufacture: embodied consequence, unanticipated interaction, persistent human response, and histories that were not authored for the learner. The Second Loop provides what reality cannot freely afford: branching, replay, controlled variation, temporal compression, and revision of the mechanisms through which later World Processes are constructed.
Their value lies in this asymmetry.
World Processes remain useful only insofar as they remain answerable to what happens outside them. A Human Agent cannot establish the preferences of a real person; a simulated interaction cannot certify behavior in a real household; internal consistency cannot substitute for external evidence. Reality remains the outer constraint on what the Second Loop is allowed to claim.
The coupling works in the other direction as well. A single lived experience need not remain a single trajectory. It can expose missing dynamics, motivate revisions to participant or environment models, and seed World Processes in which related situations can be reconstructed, varied, and examined under controlled conditions. The Second Loop therefore extends the scientific leverage of real experience without promoting internally generated histories into ground truth.
For Machine Aliveness, this makes deployment part of the experimental regime, not merely its downstream destination. Many of the phenomena we care about—persistent relationships, evolving expectations, long-horizon adaptation, and the consequences of shared history—emerge only when machines remain in the world long enough for earlier interactions to shape later ones.
Our work spans both research through Machine Aliveness and product development and deployment through Luvbotics. This continuity matters for the questions studied here: deployment exposes physical and social dynamics that designed environments cannot fully anticipate, while research turns those encounters into abstractions and World Processes that can be systematically examined.
Meet our first agent entering the world. ↗
Machine intelligence asks what a machine can do. Machine aliveness asks what it can become. The Second Loop gives that becoming another world—without severing it from the one that makes it real.
Research context
The Second Loop builds on a long-standing view of intelligence as something that develops through extended interaction rather than isolated task completion. Developmental robotics emphasizes the role of physical embodiment and interaction with both environments and other agents, while lifelong robot learning asks how experience accumulated across a machine's lifetime can shape what it learns next. [19] [20] Our concern begins from the same temporal premise, but shifts attention toward the world-side infrastructure required when development extends across persistent physical and social histories.
Simulation, world models, and open-ended learning make this infrastructure computational. Simulation separates the scale of learning experience from direct physical execution; learned world models make environmental dynamics available for prediction and imagination; systems such as POET allow environments to participate in open-ended learning. [3] [4] [11] These works turn the world from a fixed backdrop into an object that can be generated, searched, and adapted.
Persistent social agents expose another requirement. Generative Agents and Virtual Community model participants whose memories, activities, and interactions persist, while generalist embodied systems increasingly ground behavior in human demonstrations, instructions, and interactive feedback. [5] [6] [13] [14] For machines that repeatedly encounter the same people, these threads meet: human state is both supervision for the learner and part of the continuing dynamics of the world it returns to.
At the same time, the machinery surrounding intelligent systems is becoming increasingly programmable by agents themselves. Self-modifying systems such as Darwin Gödel Machine and HyperAgents make code, scaffolds, evaluators, and improvement procedures explicit objects of change, while newer agentic systems increasingly operate across software, 3D world construction, simulation, and embodied control. [1] [2] [8] [9] [15] [16] [17] [18] Related work has also argued that unexpected physical experience can require revising both an embodied actor and its computational account of the world. [12] Together, these developments suggest that agents can increasingly participate not only inside computational worlds, but in the machinery through which those worlds are constructed and revised.
The Second Loop names the systems object at this intersection. A World Process carries participants, processes, and consequences forward through time. Agents as World makes the organization of that process itself programmable. The World Kernel preserves the boundary between revisable world-making machinery and committed causal history.
The resulting perspective is not simply to place a learning agent inside a richer simulator. It is to make the continuing world—and increasingly, the machinery by which that world is made—part of the computational substrate through which a machine develops.
References
[1] Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents
Jenny Zhang et al. · Research paper · 2025.
[2] Hyperagents
Jenny Zhang et al. · Research paper · 2026.
[3] Learning dexterity
OpenAI · Research article · 2018.
[4] World Models
David Ha and Jürgen Schmidhuber · Research article · 2018.
[5] Generative Agents: Interactive Simulacra of Human Behavior
Joon Sung Park et al. · Research paper · 2023.
[6] Virtual Community: An Open World for Humans, Robots, and Society
Qinhong Zhou et al. · Research paper · 2025.
[7] SimPy: Overview
SimPy documentation · Discrete-event and real-time execution.
[8] Building effective agents
Anthropic · Engineering article · 2024.
[9] WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment
Hao Tang, Darren Key, and Kevin Ellis · Research paper · 2024.
[10] SOTOPIA-π: Interactive Learning of Socially Intelligent Language Agents
Ruiyi Wang et al. · Research paper · 2024.
[11] Paired Open-Ended Trailblazer (POET): Endlessly Generating Increasingly Complex and Diverse Learning Environments and Their Solutions
Rui Wang et al. · Research paper · 2019.
[12] Building Physical RSI Beyond the Known World
MirroS · Research position · 2026.
[13] GR00T N1.5: An Improved Open Foundation Model for Generalist Humanoid Robots
NVIDIA GEAR · Research page · 2025.
[14] Gemini Robotics
Google DeepMind · Model page · 2026.
[15] GPT-6 Astra: A new generation of intelligence
OpenAI · Product and research release · 2026.
[16] SceneSmith: AI agents create virtual playgrounds for robot training
MIT CSAIL / Toyota Research Institute · Research coverage · 2026.
[17] Agent as Policy for Robotic Manipulation
Mengzhao Jia et al. · Research paper · 2026.
[18] Claude plays robotics
Anthropic · Research article · 2026.
[19] Cognitive Developmental Robotics: A Survey
Minoru Asada et al. · IEEE Transactions on Autonomous Mental Development · 2009.
[20] Lifelong Robot Learning
Sebastian Thrun and Tom Mitchell · Robotics and Autonomous Systems · 1995.