Agent System

The agent system provides LLM orchestration patterns for domain analysis: role-based prompt execution, parallel inference coordination, temporal consolidation, and background evolution.

Terminology note: Components here are labeled “agents” following pre-2024 conventions (cf. LangChain, AutoGPT). They are orchestrated pipelines — they lack observe-reason-act loops and self-directed goal pursuit.

Swarm Execution

The primary pattern executes multiple LLM calls with distinct role-based system prompts in parallel:

Execution is parallel (asyncio.gather over parallel_inference()) but not agentic — roles do not observe each other’s outputs or iterate. Each role has lattice positioning behaviors: Leader seeks cluster centroids, Risk positions at boundaries (negative curvature regions), Adversary samples uniformly.

Latent Swarm (LatentMAS)

Reduces inter-agent token transfer by sharing embeddings instead of text via Qdrant (Guo et al., 2024). Each agent stores its output as a 768-dim Nomic embedding in the gaius_latent_memory collection. Subsequent agents retrieve relevant context via semantic search rather than receiving full text.

Token reduction: 70–90% compared to text-based coordination. The collection schema includes domain, agent_id, session_id, and timestamp — enabling both cross-agent retrieval within a session and longitudinal analysis across sessions.

ThetaAgent: Temporal Consolidation

ThetaAgent (agents/theta/agent.py) executes a five-stage deterministic pipeline for cross-temporal knowledge linking:

1. Temporal slicing — Documents organized into weekly slices (YYYY-WNN). Each slice is a consolidation unit.

2. NVAR dynamics — Nonlinear Vector AutoRegression via reservoir computing (Gauthier et al., 2021) computes a consolidation signal from embedding centroid trajectories. Given slice centroids c₁,…,cₜ ∈ ℝ⁷⁶⁸, NVAR predicts ĉₜ₊₁ and computes drift = ‖ĉₜ₊₁ − cₜ‖₂. High drift → high urgency → consolidation priority.

3. BERTSubs inference — Subsumption relationships (A ⊑ B) between concepts are inferred using BERTSubs (Chen et al., 2023) from DeepOnto. Requires an OWL domain ontology with rdfs:subClassOf axioms and JVM via JPype.

4. Knowledge Gradient selection — Candidate relationships filtered by the KG policy (Powell & Ryzhov, 2012), which balances exploration (uncertain candidates) against exploitation (high-confidence relationships). Only candidates whose expected improvement exceeds a cost threshold are selected.

5. Document augmentation — Selected relationships injected as wikilinks ([[Target]]) and action links ([action:search "query"]) into source documents.

MetaAgent Coordination

MetaAgentManager (agents/metaagent_swarm.py) coordinates specialist analysts to answer natural language questions by querying structured data sources:

• Lineage analyst — Translates questions to Cypher queries against the lineage graph
• Ops analyst — Translates to SQL queries against operational metrics
• Resource analyst — Queries infrastructure state
• Topology analyst — Queries embedding space structure

Results are synthesized by a Correlator (single LLM call). This is parallel execution with synthesis, not autonomous multi-agent reasoning.

Background Processes

Two daemons run within the engine:

• Evolution Daemon — Optimizes agent prompts during GPU idle periods (<30% utilization). Methods include APO (Zhou et al., 2023) and GEPA (Guo et al., 2024). Agent versions can be merged via TIES (Yadav et al., 2023) or DARE (Yu et al., 2024) in parameter space.
• Cognition Agent — Generates “thoughts” about patterns in KB activity (every 4–8h). Thought types: PATTERN, CONNECTION, CURIOSITY, SELF_OBSERVATION.

Subchapters

• Evolution — RLVR-based prompt optimization
• Cognition — Autonomous thought generation
• Theta Consolidation — Temporal knowledge linking
• CLT Memory — Cross-Layer Transcoder features