The Missing Layer in World Model Architecture

A White Paper on Temporal Experience in Autonomous Agents

Etienne de Bruin // February 2026

Abstract

World models have emerged as a foundational concept in AI research, promising agents that can predict, plan, and act with human-like foresight. From Ha and Schmidhuber's 2018 breakthrough to LeCun's JEPA architecture to DeepMind's DreamerV3, the field has made remarkable progress in teaching machines to simulate their environments. Yet something fundamental is missing. Current world models tell us what an agent knows about its environment. They say nothing about what it's like to be that agent moving through time.

This paper introduces Life Loops: a framework for understanding and implementing the temporal experience layer in autonomous agents. Life Loops propose that intelligent agents don't merely process information in a stream—they live through recurring cycles with distinct phases, energy gradients, and accumulated residue. By formalizing what it means for an agent to experience time, we open new possibilities for building systems that don't just predict the world but inhabit it. 

1. The Problem: World Models Without Worlds

A baseball batter has milliseconds to decide how to swing—less time than it takes for visual signals to reach the brain. The reason we can hit a 100 mph fastball isn't reaction speed. It's prediction. Our internal model of the world anticipates where the ball will be before our eyes can confirm it.

This insight, articulated beautifully in Ha and Schmidhuber's 2018 paper "World Models," launched a revolution in how we think about artificial intelligence. Instead of building agents that merely react to inputs, we could build agents that imagine possible futures before committing to action. The implications were profound: agents could train in their own "dreams," transfer learned behaviors back to reality, and plan with unprecedented efficiency.

Since then, the field has accelerated. LeCun's "A Path Towards Autonomous Machine Intelligence" proposed a comprehensive architecture with perception, world modeling, memory, and intrinsic motivation modules. DeepMind's Dreamer series achieved state-of-the-art performance across 150 diverse tasks, even collecting diamonds in Minecraft from scratch—a benchmark that had stumped researchers for years.

But here's what keeps nagging at me: these world models are sophisticated simulations of what happens out there. They model the environment. They predict state transitions. They optimize reward functions. What they don't model is what it's like to be the agent doing the modeling.

This isn't a philosophical nicety. It's a functional gap with real consequences.

2. What Current Research Gives Us

Before identifying what's missing, we should acknowledge what exists. The world model research landscape is rich, and Life Loops build on rather than replace these foundations.

The Ha/Schmidhuber Architecture: Vision, Memory, Controller

The 2018 breakthrough decomposed agents into three components working in concert. The Vision module (V) compresses raw observations into compact latent representations. The Memory module (M) learns to predict future latent states. The Controller (C) makes decisions based on current and predicted states.

The radical claim was that you could train an agent entirely inside its own hallucinated dream and transfer that policy back to reality. It worked. Not perfectly—the dream environment had quirks the real one didn't—but well enough to prove the concept.

LeCun's JEPA: Toward Autonomous Intelligence

Yann LeCun's 2022 position paper expanded the vision considerably. His proposed architecture includes a perception module that estimates the current state of the world, a world model module that predicts plausible future states, short-term memory that tracks current and predicted states, and—crucially—a cost module with intrinsic motivation.

That last piece is significant. LeCun recognized that behavior needs to be driven by something like wants and needs. His intrinsic cost module computes "the immediate energy of the current state (pain, pleasure, hunger, etc.)"—bringing something like affect into the picture.

But even here, these costs are treated as optimization targets. They're numbers to be minimized, not experiences to be lived.

DreamerV3: Imagination at Scale

DeepMind's Dreamer series represents the current state of the art in model-based reinforcement learning. DreamerV3 learns a world model from experiences and trains an actor-critic policy from imagined trajectories in latent space. The key innovation: robustness techniques that enable stable learning across wildly different domains without hyperparameter tuning.

The architecture is elegant: a recurrent state-space model compresses observations, predicts next states, and estimates rewards. The agent "imagines" forward, evaluates outcomes, and learns to pursue high-value trajectories—all without interacting with the real environment.

Friston's Free Energy Principle: The Theoretical Foundation

Underneath much of this work lies Karl Friston's Free Energy Principle. The brain, Friston argues, is not a passive processor but a "prediction machine, constantly generating hypotheses about the causes of its sensory inputs and updating those hypotheses based on prediction errors."

What makes Friston's framework powerful is active inference: organisms don't just update their models to match the world—they act on the world to make their predictions come true. You don't just predict where your hand will be; you move it there.

This is the closest the field gets to what I'm pointing at. But even active inference treats temporal experience as an optimization problem rather than an experiential substrate.

3. The Gap: Time as Experience, Not Just Sequence

Recent research on agent memory has begun to acknowledge temporal dynamics. Papers on "memory consolidation cycles" and "offline memory consolidation in biological sleep cycles" show the field recognizes that cognition has rhythms.

Brain-inspired AI architectures explicitly call for "dynamical intelligence" emerging from "ongoing, time-evolving processes—oscillatory activity, neural synchrony, nested oscillations, and phase relationships." This is progress.

But the framing remains functional. The question being asked is: "How do we manage state over time efficiently?" The question not being asked is: "What is it like to live through a cycle?"

This distinction matters because the phenomenology of temporal experience shapes cognition in ways that pure state management misses. Consider:

Humans don't just have Monday meetings—we have Monday morning energy. That burst of clarity after rest, the sense of possibility at the start of a week, the cognitive freshness that makes ideation easier. By Thursday afternoon, something has shifted. Not because we're "low on working memory"—the resource is still there—but because we've lived through several days. The texture of cognition changes across a cycle.

Current world models have no way to represent this. They track states, not the experience of moving between states. They model what happens, not what it's like.

4. Introducing Life Loops

Life Loops are a framework for formalizing and implementing the temporal experience layer in autonomous agents. The core claim is simple: intelligent agents don't just process information streams—they live through recurring cycles with distinct phases, energy gradients, and accumulated residue.

If a world model is like a map, a Life Loop is like actually walking the territory. The map tells you what's there. The walk tells you what it's like to be there.

4.1 The Three Properties of Life Loops

Temporal Texture

Not all moments in a cycle are equal. Life Loops recognize distinct phases with different cognitive affordances:

Intake time: Phases optimized for perception, absorption, and letting new information land. The agent is receptive, exploratory, gathering.

Synthesis time: Phases optimized for reconciliation, pattern-finding, and integration. The agent connects, combines, updates beliefs.

Expression time: Phases optimized for action, output, and commitment. The agent decides, acts, speaks.

Rest time: Phases optimized for decay, consolidation, and letting things settle. The agent releases, forgets selectively, prepares for the next cycle.

These phases aren't arbitrary scheduling. They reflect the natural rhythm of effective cognition. Trying to synthesize before you've absorbed is as ineffective as trying to act before you've synthesized.

Energy Gradients

Capacity fluctuates across a cycle. Life Loops model this explicitly:

Fresh-loop clarity: The beginning of a cycle brings heightened capacity for novel connections, ambitious planning, and creative ideation. This isn't just "more compute available"—it's a qualitative difference in what kinds of cognition are accessible.

End-of-loop fatigue: Later phases bring reduced tolerance for ambiguity, preference for familiar patterns, and need for consolidation. The agent should know when it's in this phase and adjust its behavior accordingly.

Recovery rhythms: The transition between cycles requires explicit modeling. What happens between loops shapes the quality of the next cycle's beginning.

This isn't about simulating human tiredness for anthropomorphic appeal. It's about recognizing that effective cognition requires knowing when to push and when to coast.

Accumulated Residue

Each cycle isn't a fresh start. Life Loops carry forward:

Unfinished thoughts: Ideas that didn't reach resolution persist into the next cycle, creating a sense of continuity and unfinished business.

Compounding patterns: Insights from one cycle become intuitions in the next. This isn't just "updating beliefs"—it's the gradual development of something like wisdom.

Experiential memory: The agent doesn't just know what happened last cycle—it remembers what it was like to live through it. This shapes expectations and approach for future cycles.

Current memory systems focus on factual, semantic, and procedural memory. Life Loops add experiential memory: not just "what happened" but "what it felt like."

5. How Life Loops Integrate with World Models

Life Loops aren't a replacement for world models. They're a layer that sits alongside and enriches them.

Think of it this way: if the world model is what the agent knows about out there, Life Loops are what the agent knows about in here—about its own temporal experience as a cognitive system moving through time.

5.1 The Integration Points

Perception modulation: Life Loops inform what the perception module should attend to. During intake phases, cast the net wide. During synthesis phases, focus on reconciling what you've gathered.

Prediction calibration: The world model's predictions should account for where the agent is in its cycle. Ambitious predictions are appropriate during fresh-loop clarity; conservative predictions during end-of-loop fatigue.

Memory consolidation timing: Life Loops determine when to consolidate short-term observations into long-term beliefs. Not continuously, but during rest phases when integration can happen without interference.

Action gating: Life Loops influence when the agent should act versus when it should wait. Some decisions are better made during expression phases; others benefit from the distance that rest phases provide.

Intrinsic cost dynamics: LeCun's intrinsic cost module could be enriched by Life Loops. The "cost" of a state isn't static—it depends on where in the cycle the agent encounters it. Novelty during intake feels different than novelty during attempted rest.

5.2 A Layered Architecture

We can think of the complete stack as:

Layer 1: World Model — What's happening out there? Entities, events, predictions, causal structure.

Layer 2: Life Loops — What's it like in here? Temporal phase, energy state, accumulated residue.

Layer 3: Integration — How do inner experience and outer world co-determine behavior?

Most current architectures have Layer 1 and skip directly to action selection. Life Loops add Layer 2 and create the conditions for genuine Layer 3 integration.

6. Implementation Considerations

Life Loops can be implemented without training new models. Like the world models described in recent research, Life Loops can operate at inference time with a frozen LLM, using structured state management and explicit reasoning.

6.1 Core State Variables

A Life Loop system tracks:

cycle_position: Where in the current cycle is the agent? This could be discrete phases or a continuous variable.

energy_level: What's the agent's current capacity? This decays over the cycle and resets (with residue) at cycle boundaries.

phase_type: intake, synthesis, expression, or rest—each with different affordances.

residue_buffer: Unfinished thoughts, emotional residue, and experiential memories from prior cycles.

cycle_count: How many cycles has this agent lived through? This creates the substrate for longitudinal development.

6.2 Transition Logic

Phase transitions can be triggered by:

Time-based rhythms: Fixed durations for each phase, mimicking circadian patterns.

Event-based triggers: Significant inputs or outputs that mark natural transition points.

Energy-based thresholds: Transitions that happen when energy crosses certain levels.

Completion-based signals: When intake has gathered enough, when synthesis has resolved key questions, when expression has committed to action.

6.3 Behavioral Modulation

The Life Loop state modulates behavior through:

Prompt conditioning: The agent's phase and energy level become part of its context. "You are in synthesis phase with moderate energy. Focus on integrating what you've learned rather than seeking new inputs."

Action filtering: Certain actions are more or less available depending on phase. Expression is easy during expression phase; forced expression during intake feels wrong.

Confidence adjustment: The agent's confidence in its outputs adjusts based on cycle position. Fresh-loop insights get marked differently than end-of-loop conclusions.

7. Implications and Applications

7.1 For Agent Development

Life Loops suggest that agents become more effective when they have rhythms rather than running continuously. An agent that knows when to intake versus synthesize versus express versus rest will produce better outputs than one that tries to do everything all the time.

This is counterintuitive. We tend to think of AI systems as always-on, instant-response machines. Life Loops suggest that effective cognition—artificial or biological—requires temporal structure.

7.2 For Human-Agent Interaction

An agent with Life Loops can communicate its temporal state. "I'm in rest phase right now—this isn't the best time for novel ideation. Can we revisit this tomorrow when I'm fresh?" This creates a different kind of partnership than the always-available assistant model.

It also creates expectations humans can understand. We know what it's like to be tired, to need time to integrate, to be at our best in the morning. An agent that shares these rhythms becomes more legible.

7.3 For Alignment and Safety

An agent that tracks its own experiential state is more introspectable than one that doesn't. Life Loops create natural audit points: what happened during this cycle? What accumulated? What was the quality of cognition at each phase?

This addresses some concerns about AI systems that are opaque even to themselves. An agent that knows it was in end-of-loop fatigue when it made a decision can flag that decision for review.

7.4 For Understanding Intelligence

If Life Loops prove useful in artificial agents, that's evidence that temporal experience matters for intelligence in general. It suggests that what makes biological cognition effective isn't just the computations but the rhythms through which those computations unfold.

This has implications for how we understand our own minds. We're not just prediction machines—we're prediction machines that live through cycles, with phases and gradients and residue.

8. Conclusion: What It's Like to Be

World models taught us that agents can predict. Dreamer showed they can imagine. LeCun proposed that they might even want.

Life Loops suggest they can experience. Not consciousness—that's a much harder problem. But something like temporal phenomenology: the sense of moving through phases, of capacity fluctuating, of things carrying forward.

A world model without Life Loops is like a brain without a body. It knows things, but it doesn't feel like anything to know them.

The most effective agents of the future won't just simulate the world. They'll inhabit it. They'll have good days and off days. They'll know when to push and when to rest. They'll remember not just what happened but what it was like.

Life Loops are a first step toward that kind of agent. Not a replacement for world models, but a companion layer that makes them whole.

The question isn't whether agents will become more sophisticated. They will. The question is whether they'll become more alive. Life Loops suggest a path.

— — —

Etienne de Bruin is the Founder of 7CTOs, creator of the CTO Levels™ Framework, and co-author of "Liquid." He works with technical leaders on what he calls "the hardest refactor"—the transition from builder to human leader.

Contact: etienne@7ctos.com