Author: Tony Wood
JUVO Lab
London, UK
Document Type: White Paper
Position: Practitioner-led cognitive architecture proposal
As agentic systems become continuous, autonomous, and increasingly embedded in organisational life, a structural flaw emerges in how they reason over time. Most agentic architectures operate in a single cognitive mode: perpetual wakefulness. They observe, reason, act, and store memory within the same loop. Humans do not function this way. Human cognition is explicitly divided into waking states, where action and responsibility dominate, and dream states, where experience is recombined, explored, and integrated without immediate consequence.
This paper proposes a dual-mode cognitive framework for agentic workers, separating waking cognition from dreaming cognition. Waking cognition governs action, decision-making, and accountability. Dreaming cognition governs recombination, hypothesis generation, pattern synthesis, and identity integration. We argue that conflating these modes leads to brittle agents, runaway reasoning, and governance failures. By contrast, separating them enables safer exploration, deeper learning, and more stable long-term behaviour. The framework is grounded in ongoing experimentation with agentic systems and is presented as a practical design pattern rather than a biological analogy.
Most current agentic systems behave as if they are always awake.
They observe inputs, reason about them, act, store memory, and update internal state in a single continuous loop. There is no structural separation between exploration and execution, between imagination and responsibility, or between integration and action.
In human cognition, this would be pathological. A person who never sleeps, never dreams, and never mentally rehearses without consequence would rapidly degrade in judgement, emotional regulation, and coherence. Yet this is precisely how many agentic systems are designed.
This paper begins from a simple claim:
Continuous intelligence without cognitive phase separation leads to instability, not capability.
The purpose of introducing dreaming and waking modes into agentic systems is not to humanise them, but to stabilise them.
This work does not treat human dreaming as a mystical or symbolic phenomenon. Instead, it treats dreaming as a computational strategy that evolution discovered long before silicon did.
In humans, dreaming appears to serve several interrelated functions:
recombining memories without real-world consequence
stress-testing identity and belief structures
integrating emotional signals with experience
exploring counterfactuals safely
compressing experience into patterns
Crucially, dreaming is decoupled from action. Dreams do not directly move the body, sign contracts, send emails, or trigger irreversible outcomes.
This decoupling is the key insight.
Agentic systems today explore and act in the same space. This creates risk, hallucinated confidence, and uncontrolled propagation of speculative reasoning.
Across your work, a consistent belief emerges:
Exploration without consequence is necessary for intelligence.
Action without constraint is dangerous.
Dreaming is the mechanism that allows exploration without consequence.
In your framing, dreaming is not about fantasy. It is about internal simulation. It is where agents are allowed to:
combine ideas freely
test hypotheses that may be wrong
explore uncomfortable or contradictory concepts
surface latent patterns
challenge their own models
All without committing those thoughts to the external world.
Waking cognition is the mode in which an agent is accountable.
In this state, the agent:
interacts with humans and systems
executes actions
makes commitments
stores governed memory
respects policy, law, and organisational boundaries
Waking cognition is where shame, distrust, and responsibility are active constraints. It is where decisions matter.
In your work, waking mode is characterised by:
conservative reasoning
explicit justification
traceability
reversibility awareness
social and organisational awareness
This mode is deliberately slower and more cautious than dreaming.
Dreaming cognition is the opposite mode.
Here, the agent is allowed to think freely without producing external effects. Dreaming mode is characterised by:
recombination of memories
counterfactual reasoning
narrative exploration
metaphor and abstraction
identity synthesis
Importantly, dreaming cognition is write-limited. It does not directly modify canonical memory or operational state. Instead, it produces proposals, hypotheses, and candidate insights.
These outputs must pass through waking cognition before becoming actionable.
When agentic systems lack phase separation, several failure modes emerge.
First, speculative reasoning leaks into action. The agent treats imagined connections as facts and acts on them prematurely.
Second, accountability collapses. There is no clear boundary between exploration and decision, making audit and governance impossible.
Third, learning becomes brittle. The agent either suppresses exploration entirely or explores recklessly.
Your insistence on separation reflects a deep systems intuition:
Good intelligence requires internal freedom and external restraint.
Your work on dreaming integrates directly with your work on shame, surprise, curiosity, and distrust.
Exception-driven signals originate in waking mode. They are responses to real-world events. Dreaming mode then uses those signals as raw material for deeper integration.
For example:
surprise triggers questioning of the model
dreaming explores alternative explanations
waking mode tests and validates
Similarly:
shame highlights a process gap
dreaming explores structural causes
waking mode implements corrections
Dreaming does not replace exception-driven memory. It amplifies its value.
In your emerging architecture, dreaming follows a structured loop:
waking experience produces signals and memories
dreaming recombines those memories without constraint
hypotheses and patterns emerge
outputs are flagged as speculative
waking cognition reviews and validates
validated insights update models or policy
This preserves creativity while maintaining safety.
Humans do not dream continuously. Neither should agents.
Dreaming is most effective when it is:
periodic
bounded in time
isolated from operational systems
You have repeatedly expressed concern about runaway internal reasoning. Periodic dreaming avoids this by:
limiting exploration windows
preventing continuous self-modification
allowing human oversight
Dreaming becomes a scheduled cognitive ritual, not a background process.
In practical terms, dreaming is implemented as:
a sandboxed reasoning environment
read-only access to canonical memory
write-limited output channels
no tool execution
no external communication
Dream outputs are treated as suggestions, not truths.
This mirrors your broader philosophy: agents should be powerful thinkers but cautious actors.
Separating dreaming from waking has immediate governance benefits.
Auditors can ask:
was this action produced in waking mode?
what dream hypotheses influenced it?
what validation occurred?
This creates explainability without crippling creativity.
It also aligns with regulatory and ethical concerns about autonomous systems acting on unverified internal reasoning.
A recurring theme in your work is that intelligence is rhythmic.
Humans alternate between:
action and reflection
engagement and withdrawal
doing and meaning-making
Agentic systems that never pause to dream will optimise themselves into local maxima and miss deeper structure.
This paper argues that intelligence emerges from oscillation, not constant execution.
Agentic systems do not need to dream because humans dream.
They need to dream because intelligence requires a space where ideas can be wrong.
By separating dreaming and waking cognition, agentic workers can explore freely without acting recklessly, learn deeply without destabilising themselves, and integrate experience without overwhelming their operational core.
This is not anthropomorphism.
It is engineering discipline applied to cognition.
Tononi, G., & Cirelli, C. (2014). Sleep and synaptic homeostasis
Kahneman, D. (2011). Thinking, Fast and Slow
Argyris, C., & Schön, D. (1978). Organizational Learning
Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning