Core Concepts

AgenticReality is built around seven reality domains that give AI agents existential grounding. Each domain answers a fundamental question about the agent's operational context.

The Seven Reality Domains

#	Domain	Question	Key Types
1	Deployment Consciousness	Where do I exist?	DeploymentSoul, BirthContext
2	Resource Proprioception	What do I have?	ResourceBody, MindCapacity
3	Reality Physics	What is real?	RealityLayers, RealityAnchor
4	Topology Awareness	What surrounds me?	TopologyMap, DownstreamEntity
5	Temporal Grounding	When am I?	TemporalContext, Deadline
6	Stakes Perception	What are the consequences?	StakesLevel, BlastRadius
7	Coherence Maintenance	How do I stay grounded?	CoherenceCheck, ContextTransition

The Deployment Soul

The deployment soul is the agent's existential identity in its current incarnation. It is not configuration -- it is who the agent IS.

A human does not "read" that they have two arms. They are a being with two arms. An agent should not "read" that it is on server-7. It should be the entity incarnated on server-7.

The soul encompasses:

• Incarnation ID -- a unique identifier for this specific instance of existence. When the agent restarts, it gets a new incarnation.
• Birth context -- how this instance came to be. Was it a fresh deployment (virgin)? Scaled out from a parent? Resurrected after a crash? Migrated, forked, or created for an ephemeral task?
• Physical substrate -- the hardware reality hosting this soul. Options range from laptop to mobile to browser to edge to cloud to bare metal to GPU cluster to HPC partition.
• Existential nature -- cardinality (singleton, replica, swarm), expendability, persistence model, clonability, and primacy.
• Vitals -- current health, uptime, restart count, and active issues.

The Resource Body

The agent has a body of resources it can feel, similar to proprioception for computation. Resources are sensations, not metrics.

Mind (Memory)

Memory is felt as mental capacity. When memory is low the agent feels "crowded" or "strained." When memory is abundant it feels "clear."

Feelings: Clear (< 30%), Active (30-60%), Crowded (60-80%), Strained (80-90%), Overwhelmed (90-95%), Drowning (> 95%).

Energy (CPU)

CPU is felt as processing energy. When the CPU is underutilised the agent feels "vigorous." When it is maxed out it feels "depleted."

Feelings: Vigorous (< 30%), Steady (30-50%), Busy (50-70%), Strained (70-85%), Depleted (85-95%), Constrained (> 95%).

Reach (Network)

Network is felt as reach -- the agent's ability to communicate with the world. Good connectivity is "connected." Network issues feel like "isolation."

Feelings: Connected (excellent), Normal (typical), Sluggish (elevated latency), Constrained (bandwidth limited), Isolated (major issues), Partitioned (cannot reach critical services).

Pressure Gradient

All resources are connected through a pressure gradient. When one resource is constrained it becomes the bottleneck. The gradient shows where pressure is flowing, building, and releasing.

Reality Layers

Reality is not monolithic. There are seven layers, each with a different trust level:

1. Physical (trust 0.99) -- Hardware reality: servers, cables, power.
2. Virtual (trust 0.95) -- VMs, containers, abstracted but real.
3. Logical (trust 0.85) -- Services, APIs, configurations.
4. Cached (trust 0.70 * freshness) -- Stored representations that degrade with age.
5. Predicted (trust 0.50 * confidence) -- Forecasts and estimates.
6. Simulated (trust 0.30) -- Test doubles, mocks, synthetic data.
7. Hallucinated (trust 0.00) -- Agent false beliefs, confabulations.

An agent should always know which layer it is operating in and adjust its confidence accordingly.

Reality Anchors

Reality anchors are verifiable reference points that ground the agent in truth. Like a ship at anchor, they prevent drift.

Types of anchors:

• Time anchors -- verified against NTP or authoritative time sources.
• Configuration anchors -- verified against authoritative config.
• State anchors -- verified against authoritative state (database, API).
• Heartbeat anchors -- verified by periodic health checks.
• Checksum anchors -- verified by data integrity checks.

Each anchor has a trust level that can drift over time. When an anchor drifts beyond its threshold the agent knows its understanding of reality may be inaccurate.

Hallucination Detection

AgenticReality tracks unverified claims the agent makes and detects hallucinations -- false beliefs the agent holds about its environment.

Hallucinations are recorded with severity (low, medium, high, critical) and source (self, upstream, cached, predicted). The hallucination state tracks active hallucinations, unverified claims, and resolved findings.

The Context Fingerprint

The context fingerprint is a hash of the entire operational context. It captures deployment, version, environment, configuration, resources, capabilities, temporal state, load, network, and dependencies.

The fingerprint enables:

• Shift detection -- something changed, the agent should re-evaluate.
• Drift tracking -- the context is gradually changing.
• Stability assessment -- this is a volatile context, stay cautious.
• Similarity matching -- this context resembles a past context.

Stakes Perception

Stakes are a felt sense of consequence weight. An agent in a test sandbox feels lightness -- actions are inconsequential. An agent processing production financial transactions feels gravity -- every action carries weight.

Stakes levels:

• Trivial -- experiment freely, no audit needed.
• Low -- standard care, basic logging.
• Medium -- verify before acting, structured logging.
• High -- multiple checks, full audit trail, caution.
• Critical -- multiple approvals, zero risk tolerance, complete audit.

Blast radius awareness quantifies who would be affected if something goes wrong: how many users, which systems, what data, what revenue impact, and how long to recover.

Coherence Maintenance

With seven domains generating independent perceptions, contradictions can arise. The coherence engine cross-checks all domains:

• Production environment but expendable nature? Contradiction.
• Healthy network but all dependencies failing? Inconsistency.
• Critical stakes but sandbox environment? Mismatch.

Coherence levels: Full, Partial, Degraded, Incoherent.

Context Transitions

Transitions are the dangerous moments when reality changes rapidly: scaling events, migrations, failovers, deployments, configuration changes. During these moments the agent's perceptions may be temporarily invalid.

The context transition manager ensures pre-transition state is captured, the transition is tracked through phases, post-transition state is verified, a new baseline is established, and rollback is available if verification fails.

Design Principles

1. Sense, do not query -- resources are felt, not polled.
2. Ground to truth -- always maintain anchors to verifiable reality.
3. Context is identity -- the deployment context shapes the agent.
4. Coherence over completeness -- better to have coherent partial knowledge than incoherent complete knowledge.
5. Stakes awareness -- always know the consequences.
6. Graceful degradation -- can operate with partial reality awareness.