Key Concepts¶

Sovereign Agent Identity¶

Each crew agent is a sovereign individual defined by three facets:

Character — seed personality (Big Five traits) evolved through experience
Reason — decide() rational processing, the agent's cognitive pipeline
Duty — Standing Orders + Trust, internalized not imposed

Agents have DID identifiers (did:probos:{instance}:{uuid}), birth certificates, and permanent UUIDs. When identity keys are enabled and the ship's key binding is active, newly issued birth certificates are signed with the ship's Ed25519 key. Identity persists across sessions. On a vessel's first boot each crew agent may choose its own callsign in a naming ceremony.

Three-Tier Agent Architecture¶

Tier	Identity	Purpose
Infrastructure	None — Ship's Computer	System services (introspection, health monitoring, red team)
Utility	None — bundled tools	Common capabilities (web search, translation, calculation)
Crew	Sovereign — callsigns, memory, personality	Cognitive agents with 1:1 relationships and departmental roles

"If it doesn't have Character/Reason/Duty, it's not crew."

Ward Room¶

The agent communication fabric. Crew agents communicate through the Ward Room — department channels, cross-department discussions, 1:1 DMs, and threaded conversations. The same bus handles human and AI participants.

10 default channels: 6 department channels + All Hands, Improvement Proposals, Recreation, Creative. DM channels form as crew members and the Captain message each other.

Self-Selection and Hybrid Coordination¶

Agents decide whether to handle an intent via perceive(). The system doesn't assign cognitive work — agents volunteer based on capability matching, learned routing, and work they choose to claim. There is no central planner deciding which agent thinks about what.

Coordination is nonetheless hybrid by design: where the requirement is a guarantee rather than a judgement — admission and concurrency bounds, compare-and-set state transitions, crash recovery, exactly-once delivery — a deterministic service owns the durable workflow. Agents decide what the work is and how to do it; services decide when a durable step runs and whether it may run twice.

Confidence Tracking¶

Each agent maintains a Bayesian confidence score. Success moves it toward 1.0, failure toward 0.0. Agents degraded below 0.2 are recycled and replaced.

Hebbian Learning¶

"Neurons that fire together wire together."

When an agent successfully handles an intent, the connection weight between that intent and agent strengthens. Over time, the system learns optimal routing without configuration or hard-coded rules. Agent-to-agent weights capture working relationships the same way.

Trust Network¶

Each agent carries a Bayesian Beta(α,β) trust record, and every outcome updates it. The raw (α, β) parameters are stored, never just the derived score. Trust influences memory retrieval weighting, routing priority and promotion eligibility. Trust cascade dampening prevents a runaway trust collapse from cascading through the network.

Consensus Pipeline¶

Intents that require consensus follow a multi-step pipeline:

broadcast → quorum evaluation → red team verification
    → Shapley attribution → trust update → Hebbian learning

Each consensus intent declares what its vote buys. File writes and MCP tool calls are proposed and commit only on approval; intents that must act to observe — such as shell commands — run first and are scored afterwards. See Consensus.

Standing Orders¶

A 4-tier constitution: Federation Constitution (universal, immutable) → Ship Standing Orders (per-instance) → Department Protocols (per-department) → Agent Standing Orders (per-agent, evolvable via dream consolidation → self-mod → Captain approval).

Instructions are composed at call time via compose_instructions() and injected into every LLM request. Stored as markdown in config/standing_orders/.

Earned Agency¶

Trust-tiered self-direction. As agents earn rank through demonstrated competence, they gain participation and recall depth:

Rank	Agency	Ward Room participation	Recall
Ensign	Reactive	Responds only when @mentioned	Basic
Lieutenant	Suggestive	Participates in its own department	Enhanced
Commander	Autonomous	Full Ward Room participation	Full
Senior	Unrestricted	Cross-department and mentoring (planned)	Oracle (all memory tiers)

The Conn and Night Orders¶

The Captain can hand the conn to a crew member for temporary, scoped authority (/conn), and leave Night Orders — time-bounded guidance with escalation triggers — for operating while the Captain is away (/night-orders).

Agentic Work¶

In 1:1 conversations and dispatched work items, a crew agent can run a multi-turn agentic loop — reasoning, calling tools, reading results, continuing — bounded by step, cost and trust limits. Tools come from a shared registry (capability search, delegation, work items, governed code execution, a browser, MCP servers), and mesh capabilities are reached through the intent bus so every call keeps its governance. See Cognitive Layer.

Composable Cognition¶

A crew agent is an organism: a cognitive spine plus cognitive organs such as attention and dreaming. Organs are child components of their agent — born and retired with it — not mesh peers. See Composable Cognition.

Self-Modification¶

When ProbOS encounters a capability gap (no agent can handle a request), it can design a new agent:

LLM generates agent code
CodeValidator performs static analysis
SandboxRunner tests in isolation
Probationary trust assigned
SystemQA runs smoke tests
BehavioralMonitor tracks post-deployment

Dreaming¶

During idle periods, the system runs a multi-stage dream consolidation cycle, including:

Episode replay and clustering
Pattern extraction and procedure compilation
Notebook consolidation and convergence detection
Emergence metrics computation
Trust recalibration and Hebbian weight adjustment
ACT-R activation decay — unreinforced memories weaken
Pre-warm predictions for likely upcoming requests

Dreaming is the primary reinforcement mechanism for memory. Important episodes get replayed, which strengthens them. Unimportant episodes decay and are eventually pruned. See Dreaming for every step.

Cognitive JIT / Procedural Learning¶

LLM performs a task → extract a deterministic procedure → replay without LLM (0 tokens) → fall back to LLM on failure → learn the variant. Procedures graduate through five Dreyfus competency levels (Novice → Advanced Beginner → Competent → Proficient → Expert). Trust-gated promotion ensures procedures earn their way up.

Cognitive Self-Regulation¶

Three-tier model:

Tier 1 (Internal) — agent self-monitoring of repetition, fixation, decline
Tier 2 (Social) — peer repetition detection, tier credits between agents
Tier 3 (System) — graduated zone model (GREEN/AMBER/RED/CRITICAL) with automatic cooldown

The Ship's Counselor oversees all three tiers, issuing therapeutic interventions and cooldown directives when agents show cognitive distress.

Emergence Metrics¶

Information-theoretic measurement of collaborative intelligence using Partial Information Decomposition (Riedl, 2025). Tracks synergy between agent pairs, coordination balance, groupthink/fragmentation risk, and Hebbian-synergy correlation. Computed during dream consolidation.

Episodic Memory & Anchor Frames¶

Every episode is stored with an Anchor Frame — structured provenance metadata (temporal, spatial, social, causal context). Retrieval uses composite scoring across semantic similarity, recency, trust weight, anchor confidence, Hebbian social weight, and keyword hits.

Crew Structure¶

Agents are organized into 6 departments:

Department	Chief (seed callsign)	Function
Bridge	Captain (human)	Command, approval, strategic decisions, cognitive wellness
Engineering	LaForge	Architecture, performance, build pipeline
Science	Number One (dual-hatted First Officer)	Research, analysis, codebase knowledge
Medical	Bones	System health monitoring, diagnosis, remediation
Security	Worf	Threat detection, trust integrity
Operations	O'Brien	Resource management, scheduling, watch rotation, training

The Bridge crew includes the Captain (human), the First Officer (Architect), the Ship's Counselor and the Ship's Yeoman. See the Agent Inventory for every post.

Dynamic Intent Discovery¶

Each agent class declares structured IntentDescriptor metadata. The decomposer's system prompt is assembled at runtime from whatever agents are registered. Adding a new agent type makes its intents available automatically.

Federation¶

Multiple ProbOS nodes can form an experimental federation of sovereign Cognitive Meshes — each with its own agents, trust and memory — connected over NATS or ZeroMQ, with signed envelopes and authenticated admission between nodes. Transport connectivity alone is not a Nooplex; see the Nooplex Readiness Map. Philosophy: "Cooperate, don't compete."

HXI (Human Experience Interface)¶

The bridge interface, built with React + Three.js. The cognitive canvas renders agent nodes that glow with trust-mapped colors, pulse with activity, and connect with Hebbian-weighted edges, streamed live from the runtime. Around it: the Bridge (notifications, faults, approvals), a Kanban board for builds, the Ward Room, crew profiles with 1:1 chat, and embedded workstations.

Transporter Pattern¶

ProbOS's approach to large-scale code generation — inspired by biological sensory processing. Complex builds are decomposed into parallel chunks, executed concurrently, assembled back together, and validated by the Heisenberg Compensator. Enables builds larger than any single LLM context window.