Agent Utilities — Concept Overview¶
See docs/concept_map.md for the canonical concept registry. See docs/pillars/architecture_c4.md for C4 architecture diagrams.
Pillar Summaries¶
- Pillar 1: Graph Orchestration Engine
- Pillar 2: Epistemic Knowledge Graph
- Pillar 3: Agentic Harness Engineering
- Pillar 4: Ecosystem & Peripherals
- Pillar 5: Agent OS Infrastructure
Engine Facades¶
| Engine | Concept | Path | Description |
|---|---|---|---|
| IntelligenceGraphEngine | KG-2.0 | knowledge_graph/core/engine.py |
Core graph engine with 15-phase pipeline |
| TopologicalAnalysisEngine | KG-2.5 | knowledge_graph/core/topological_analysis_engine.py |
Analogy, spectral, blast radius |
| EvaluationEngine | AHE-3.1 | harness/evaluation_engine.py |
Decomposed reward signals + trace distillation |
| ParallelEngine | ORCH-1.8 | graph/parallel_engine.py |
Unified 1→300+ agent parallel execution engine |
| Gateway Aggregator | AU-OS.config.gateway-service-dashboard | gateway/aggregator.py |
50-widget parallel service dashboard data layer |
The 5 Core Pillars Architecture¶
graph TD
%% Pillar 1: Graph Orchestration Engine
subgraph P1 ["Pillar 1: Graph Orchestration Engine"]
ORCH10["<b>ORCH-1.0: Core Orchestration Engine</b>"]
ORCH11["<b>ORCH-1.1: HTN Planning Pipeline</b>"]
ORCH12["<b>ORCH-1.2: Specialist Routing</b>"]
ORCH13["<b>ORCH-1.3: Execution Safety</b>"]
ORCH14["<b>ORCH-1.4: Capability Wiring</b>"]
ORCH16["<b>AU-ORCH.planning.legal-automation-roadmap: DSTDD Pipeline</b>"]
ORCH17["<b>AU-ORCH.planning.spec-driven-pipeline: Prediction Linkage Layer</b>"]
ORCH18["<b>AU-ORCH.planning.journey-milestone: RecursiveMAS Latent Orchestrator</b>"]
ORCH125["<b>ORCH-1.8: Parallel Execution & Synthesis Engine</b>"]
ORCH127["<b>AU-ORCH.execution.autonomous-department-orchestration: Dept Orchestration</b>"]
ORCH128["<b>ORCH-1.10: Reactive Dispatch</b>"]
ORCH129["<b>AU-ORCH.sandbox.compiled-orchestration-kernel: WASM Sandbox</b>"]
ORCH141["<b>ORCH-1.41-1.43: Ontology-to-Workflow Execution</b>"]
ORCH145["<b>ORCH-1.45: Queue-Driven Agent Dispatch</b>"]
end
%% Pillar 2: Epistemic Knowledge Graph
subgraph P2 ["Pillar 2: Epistemic Knowledge Graph"]
KG20["<b>KG-2.0: Active Knowledge Graph</b>"]
KG21["<b>KG-2.1: Tiered Memory</b>"]
KG22["<b>KG-2.2: Ontology & Epistemics</b>"]
KG23["<b>KG-2.3: Unified Retrieval & Graph Integrity</b>"]
KG24["<b>KG-2.4: Inductive Hypergraphs</b>"]
KG25["<b>KG-2.5: Topological Analysis</b>"]
KG26["<b>KG-2.6: Domain Ontologies & Vertical Subgraphs</b>"]
KG28["<b>KG-2.6: Memory Stability</b>"]
KG29["<b>KG-2.7: Quant Orchestration</b>"]
KG215["<b>KG-2.7: Transaction Proxy</b>"]
KG216["<b>KG-2.7: Rust-Native High-Performance Compute (FFI)</b>"]
KG219["<b>KG-2.7: Event Backbone</b>"]
KG220["<b>KG-2.7: Query Router</b>"]
KG221["<b>KG-2.21: Working Set Manager</b>"]
KG260["<b>KG-2.7: Single Company Brain</b>"]
KG255["<b>KG-2.55-2.57: Kafka Ingest Scale-Out</b>"]
KG258["<b>AU-KG.sharding.tenant-partitioned-sharding-hrw: Tenant-Sharded Engines (HRW)</b>"]
end
%% Pillar 3: Agentic Harness Engineering
subgraph P3 ["Pillar 3: Agentic Harness Engineering"]
AHE30["<b>AHE-3.0: Harness Core</b>"]
AHE31["<b>AHE-3.1: Evaluation Engine</b>"]
AHE32["<b>AHE-3.2: Evolution Engine</b>"]
AHE33["<b>AHE-3.3: Team Optimization</b>"]
AHE34["<b>AHE-3.4: Distributed Evolution</b>"]
AHE35["<b>AU-AHE.harness.self-evolution-narrative: Heavy Thinking</b>"]
AHE36["<b>AU-AHE.harness.evolution-checkpoint: Backtest & Curriculum</b>"]
AHE315["<b>AU-AHE.harness.self-improvement-overview: Interpretability & Model Evolution</b>"]
AHE321["<b>AHE-3.18-3.21: Failure-Driven Evolution & Branch Publication</b>"]
end
%% Pillar 4: Ecosystem & Peripherals
subgraph P4 ["Pillar 4: Ecosystem & Peripherals"]
ECO40["<b>ECO-4.0: Tool Interface & MCP</b>"]
ECO41["<b>ECO-4.1: A2A Network</b>"]
ECO42["<b>AU-ECO.toolkit.journey-map-narrative: Telemetry & Ecosystem</b>"]
ECO43["<b>AU-ECO.ui.company-infrastructure-orchestration: Market Data</b>"]
ECO44["<b>AU-ECO.toolkit.journey-map-adoption: KG MCP Server</b>"]
ECO410["<b>AU-ECO.mcp.toolkit-live-discovery: Dynamic Capability Ingestion & Discovery</b>"]
ECO414["<b>ECO-4.14: Infrastructure Blueprint Library</b>"]
ECO415["<b>AU-ECO.bus.pluggable-queue-backend: Queue Backend</b>"]
ECO416["<b>AU-KG.memory.team-startup-context: Automated Documentation & AGENTS.md Governance</b>"]
ECO418["<b>AU-OS.governance.lint-enforcement-hook: Lint Enforcement</b>"]
ECO419["<b>AU-ECO.toolkit.self-documenting-plugin-bundle: Plugin Bundles</b>"]
ECO420["<b>AU-OS.governance.permission-policy: Ecosystem Governance & Policy Engine</b>"]
ECO434["<b>AU-ECO.mcp.profile-differences-from-client: Multiplexer Child Resilience</b>"]
end
%% Pillar 5: Agent OS Infrastructure
subgraph P5 ["Pillar 5: Agent OS Infrastructure"]
OS50["<b>OS-5.0: Agent OS Kernel</b>"]
OS51["<b>OS-5.1: Security & Auth</b>"]
OS52["<b>OS-5.2: Resource Scheduling</b>"]
OS53["<b>AU-OS.governance.reactive-multi-axis-budget: OS Guardrails & Safety Boundaries</b>"]
OS54["<b>AU-OS.governance.wasm-micro-agent-sandbox: Telemetry</b>"]
OS56["<b>OS-5.5: Massive Scale</b>"]
OS514["<b>OS-5.14: Server-Minted JWT Identity</b>"]
OS516["<b>OS-5.16-5.18: Externalized Durable State</b>"]
OS523["<b>AU-OS.observability.no-op-without-metrics: Gateway Hardening & /metrics</b>"]
OS524["<b>OS-5.24-5.29: Fleet Autonomy Control Plane</b>"]
end
%% Cross-pillar relationships
ORCH10 <--> KG20
ORCH11 --> ORCH12
ORCH12 --> KG22
ORCH14 --> ECO40
KG21 --> KG20
KG25 --> KG20
KG26 --> KG20
KG29 --> KG20
KG215 --> KG20
KG216 --> KG20
KG219 --> KG20
KG220 --> KG20
KG221 --> KG216
KG260 --> KG22
AHE31 --> KG20
AHE33 --> ORCH12
AHE34 --> ECO42
ECO44 --> KG20
ECO410 --> ECO40
ECO415 --> ECO44
OS51 --> ORCH13
OS53 --> OS51
OS54 --> KG20
OS56 --> OS50
ECO416 --> KG20
ECO420 --> OS51
ORCH145 --> OS516
KG255 --> KG20
KG258 --> KG20
AHE321 --> OS524
ECO434 --> OS523
OS514 --> OS51
OS524 --> OS514
style P1 fill:#dae8fe,stroke:#6c8ebf,stroke-width:2px
style P2 fill:#d5e8d4,stroke:#82b366,stroke-width:2px
style P3 fill:#fff2cc,stroke:#d6b656,stroke-width:2px
style P4 fill:#e6ccff,stroke:#9673a6,stroke-width:2px
style P5 fill:#cce5ff,stroke:#004085,stroke-width:2pxConcept Index¶
Canonical Registry: See concept_map.md for the full canonical concept registry with module paths.
Pillar 1: Graph Orchestration Engine (ORCH-1.0 – 1.45)¶
| ID | Concept | Description |
|---|---|---|
| ORCH-1.0 | Core Orchestration Engine | Pydantic Graph-based DAG execution with state management and multi-agent execution |
| ORCH-1.1 | HTN Planning Pipeline | Recursive hierarchical task network decomposition |
| ORCH-1.2 | Specialist Routing & Discovery | Ontological routing, specialist tag loading, and fallback chains |
| ORCH-1.3 | Execution Safety & State | Checkpointing, retry, and state persistence |
| ORCH-1.4 | Capability Wiring Engine | Dynamic capability discovery and capability auto-activation |
| AU-ORCH.planning.legal-automation-roadmap | DSTDD Pipeline | Design-Spec-Test Driven Development lifecycle |
| AU-ORCH.planning.spec-driven-pipeline | Prediction Linkage Layer 🔬 | Prediction linking across execution iterations |
| AU-ORCH.planning.journey-milestone | RecursiveMAS Latent Orchestrator 🔬 | Continuous latent space multi-agent recursion and projection |
| ORCH-1.8 | Parallel Execution & Synthesis Engine | Unified 1→300+ agent execution engine with concurrency, DAG scheduling, and output synthesis |
| AU-ORCH.execution.autonomous-department-orchestration | Autonomous Department Orchestration | OWL-materialized company departments with reportsTo hierarchy |
| ORCH-1.10 | Reactive Event Sourcing | Reactive event-driven state and graph staging dispatcher |
| AU-ORCH.sandbox.compiled-orchestration-kernel | WASM Micro-Agent Sandbox | Isolated micro-agent WebAssembly sandbox runner with gas/memory limits and Python emulation fallback |
| ORCH-1.27 | Role-Specialized Model Routing | Binds planner/generator/learner/judge + RLM (executor/proposer/sub-LM) roles to model tiers+tags over the registry |
| ORCH-1.28 | Composable Skills + Generic Adapter | Structured Skill units (instructions+packages+modules+tools) + merge; minimal generic env adapter preserving the host evaluator |
| ORCH-1.29 | RLM Resilience + Telemetry | Structured RunTrace + FailureClass taxonomy; recoverable tool timeout vs fatal sandbox error |
| ORCH-1.30 | Generalizing GEPA | Held-out feedback/Pareto split + AgentSpec anti-overfit grounding + held-out candidate selection (transferable skills) |
| ORCH-1.31 | Graph-Native Optimization State | Persist the GEPA Pareto frontier + ancestry to the epistemic-graph; resumable, cross-session optimization |
| ORCH-1.41 | Process Plan Compiler | compile_process MCP/REST: lifts a descriptive BPMN process into an executable plan (ProcessPlanCompiler) |
| AU-ORCH.execution.ontology-validation-execution-path | Execution Ontology Gate | Ontology validation on the execution path (workflow_gate.py) before a compiled process runs |
| ORCH-1.43 | Workflow Lineage Close-Out | Run lineage written back to the KG, closing the descriptive↔executable provenance loop |
| AU-ORCH.session.durable-goal-registry-goals | Durable Goal Registry | Goals persist across restarts; stranded runs rehydrate as orphaned instead of silently vanishing |
| ORCH-1.45 | Queue-Driven Agent Dispatch | Session-keyed agent_turns queue (AgentTurnEnvelope) + stateless agent-dispatch-worker fleet with fleet-visible placement |
| AU-ORCH.routing.resolve-body-single-canonical | Canonical StructuredPrompt | One StructuredPrompt shape + single canonical body resolver; per-package prompts compose onto the agent-utilities base via extends (authored with the prompt-builder skill) |
Pillar 2: Epistemic Knowledge Graph (KG-2.0 – 2.58)¶
| ID | Concept | Description |
|---|---|---|
| KG-2.0 | Active Knowledge Graph | Core 15-phase pipeline, OGM, IntelligenceGraphEngine |
| KG-2.1 | Tiered Memory & Context 🔬 | Episodic/semantic/procedural memory, context compaction |
| KG-2.2 | Ontology & Epistemics | OWL ontology bridge, FIBO/BFO, semantic subsumption |
| KG-2.3 | Unified Retrieval & Graph Integrity 🔬 | Fingerprinting, vectorized semantic indexing, hybrid retriever, consistency validation |
| KG-2.4 | Inductive Knowledge | Knowledge synthesis and cross-pillar synergy engine |
| KG-2.5 | Topological Analysis | Analogy engine, spectral clusters, blast radius |
| KG-2.6 | Domain Ontologies & Vertical Subgraphs | Aggregated vertical domains including Finance, Enterprise, Company Operations, and Research |
| KG-2.6 | Memory Stability | Self-reflecting memory observer and stability checks |
| KG-2.7 | Multi-Domain Architecture | Decoupled graph frameworks and multi-domain graph orchestration |
| KG-2.7 | Transaction Proxy | Centralized gateway and transactional persistence layer |
| KG-2.7 | Rust-Native High-Performance Compute | High-performance quantitative execution, graph traversal, and epistemic reasoning via the out-of-process epistemic-graph engine (MessagePack/UDS client; no PyO3) plus Rustworkx |
| KG-2.7 | Event Backbone | Protocol-based pub/sub with MemoryEventBackend (default) and RedpandaEventBackend (distributed) |
| KG-2.7 | Query Router | Cost-based query planner routing reads across the engine authority's compute, cache, persistent, and vector paths |
| KG-2.11 | Bi-Temporal Memory Layers | Event-time vs storage-time + valid_from/valid_to on the graph; as-of queries and event-time contradiction precedence; procedural memory layer |
| KG-2.12 | Memory-First Retrieval (HyDE) | HyDE query expansion + dual thresholds + self-correcting two-pass + quantitative-fidelity ledger over the hybrid retriever |
| KG-2.13 | Background Learning Engine | Async, semaphore-bounded learner emitting typed, outcome-grounded ADD/UPDATE/DELETE bi-temporal memory edits |
| AU-KG.memory.ground-truth-preamble-declaring | Ground-Truth Context Authority | Authority-ranked startup context + a Ground-Truth preamble so injected memory is treated as authoritative (no re-fetching) |
| KG-2.15 | Resilient Retrieval | 4-level retrieval fallback cascade + social-closer triviality gate |
| KG-2.17 | Memory Hygiene | Scheduled decay scanner (archive via valid_to, never delete) + semantic-merge dedup |
| KG-2.18 | Evidence-Weighted Memory | Bayesian trust feedback loop + recall/usage telemetry + generation lineage extending the quality gate |
| KG-2.19 | Self-Curating Wiki | Delta-skip (SHA-256) continuous ingest of a markdown knowledge vault, reusing the ingestion engine + synthesis |
| KG-2.20 | Rust-Native Finance Compute Suite | epistemic-graph quant kernels (EG-KG.domains.market-microstructure-sizing-backtest/g/h/i): market-making (Avellaneda-Stoikov/GLT/logit), microstructure (OFI/VPIN/microprice/Hawkes), sizing (Kelly/Bayesian/empirical), validation (purged-CPCV/DSR/PBO/Brier), forensic scores, state-space/stat-arb (Kalman/OU/ADF), signal combination (alpha-engine/IR=IC√N) |
| KG-2.21 | Working Set Manager | LRU-evicting subgraph cache for the Rust engine authority with 50K node cap |
| KG-2.22 | Data Science Primitives | Rust-backed OLS / K-means / PCA / estimators (ridge/lasso/RF/GB/SVR) replacing scikit-learn on the hot path, parity-validated |
| KG-2.7 | Single Company Brain | Extensible operational state layer encompassing Ontology Bridges, Enterprise Architecture Repositories, and Entailment-Aware Permissions |
| AU-KG.ontology.populated-at-import-real-3 | Remote VCS Enumeration | Enterprise-scale ingestion: enumerate every repository across a GitHub org/user or GitLab instance/groups (keyset / affiliation pagination) into a manifest for bulk workspace onboarding (repository-manager vcs_enumerator) |
| KG-2.52 | Ontology Publisher Tick | Background publish of the authoritative TBox to Fuseki (core/ontology_publisher.py) |
| AU-KG.ontology.descriptive-process-world-gains | BPMN Process Lift | Step-level shape for the descriptive process world (Camunda extractor + owl_bridge) |
| AU-KG.ingest.cross-host-safe-kg | Cross-Host Task Queue | Atomic SKIP LOCKED claims + visibility-timeout recovery on the shared Postgres state store |
| KG-2.55 | Fail-Loud Queue Backend Selection | TASK_QUEUE_BACKEND=sqlite\|postgres\|kafka; explicit backends fail loud at startup instead of silently degrading |
| KG-2.56 | Keyed Ingest Partitions | kg_tasks partition keys (tenant → repo/corpus → task type) for per-tenant/per-repo ordering |
| AU-KG.ingest.decoupled-kg-ingest-consumer | Decoupled kg-ingest Consumer Group | kg-ingest-worker runs ingest as engine clients on any host; at-least-once, idempotent claims, lag metrics |
| AU-KG.sharding.tenant-partitioned-sharding-hrw | Tenant-Partitioned Engine Sharding | HRW graph→shard routing over GRAPH_SERVICE_ENDPOINTS with tenant→named-graph placement |
| KG-2.65 | GPU-Slot-Scheduled Fact Extraction | A single shared GPU slot serializes LLM fact-extraction so long training/ingest runs cooperate (knowledge_graph/ingestion/gpu_slot_scheduler.py); data-science-mcp training jobs share the same slot |
| KG-2.70 | Evidence-Subgraph Task Synthesis | Build a bounded evidence-graph workspace around an answer entity for shortcut-resistant deep-search task synthesis (knowledge_graph/search_synthesis/evidence_subgraph.py; distills FORT-Searcher arXiv:2606.12087) |
| AU-KG.retrieval.formulate-adversarially-refine | Shortcut-Risk Detectors | Four graph-query detectors — single-clue selectivity, evidence co-coverage, exposed constants, prior-knowledge binding — over the evidence graph (knowledge_graph/search_synthesis/shortcut_risks.py) |
| AU-KG.retrieval.question-formulation-adversarial-refinement | Question Formulation & Adversarial Refinement | Render a clue subgraph as a verifiable question (name withholding) and adversarially refine until no shortcut trips (knowledge_graph/search_synthesis/question_formulation.py) |
| AU-KG.compute.kg-3 | Learned World-Model Backend + SAI Track | Parametric latent-dynamics backend for the world model that generalizes to unseen (state, action) (ridge map over embeddings), plus a WorldModelVerifier making prediction accuracy a SAI specialization domain (knowledge_graph/core/world_model.py, harness/world_model_task.py) |
| KG-2.9 | Vendor-Neutral Enterprise Crosswalk | Bidirectional source connectors with each per-system class related to one canonical concept, making cross-vendor enterprise reasoning vendor-neutral (ServiceNow/ERPNext, LeanIX/Camunda/ARIS); risk-tier approval queue for high-stakes write-backs |
| AU-KG.ontology.kyle-insider-stealth-surveillance | Finance Microstructure / Kyle-Surveillance Ontology | MicrostructureSignal and SurveillanceSignal OWL interfaces plus typed links (grounded_in Article / relates_to Concept) and promoted node types for insider/stealth surveillance signals (knowledge_graph/ontology/finance_objects.py) |
| AU-KG.retrieval.pauserec-implicit-reasoning-generative | PauseRec Implicit-Reasoning Generative Recommender | Inference-time deterministic latent-refinement adaptation in core (retrieval/generative_recommender.py, torch-free); the trainable-<pause>-token TRAINING track is re-homed to data-science-mcp (data_science_mcp/training/pause_token_trainer.py) per the dependency-discipline rule |
Pillar 3: Agentic Harness Engineering (AHE-3.0 – 3.21)¶
| ID | Concept | Description |
|---|---|---|
| AHE-3.0 | Agentic Harness Core | Harness lifecycle, initialization, SDD integration |
| AHE-3.1 | Continuous Evaluation Engine | Multi-strategy EvalRunner, decomposed rewards |
| AHE-3.2 | Agentic Evolution Engine | Skill neologism, config versioning, variant pool |
| AHE-3.3 | Team & Synergy Optimization | TeamConfig, coalition composition, synergy scoring |
| AHE-3.4 | Distributed Agentic Evolution | Self-model, stability, ecosystem PR generation |
| AU-AHE.harness.self-evolution-narrative | Heavy Thinking & Background Intelligence | Heavy thinking, background intelligence |
| AU-AHE.harness.evolution-checkpoint | Backtest & Curriculum | Backtest harness, horizon-aware curriculum |
| AU-AHE.harness.self-improvement-overview | Interpretability & Model Evolution | Agent-Interpretable Model Evolver workflows and LLM-Graded Interpretability Tests |
| AHE-3.12 | LongMemEval-S Validation Harness | FastAPI /benchmark surface (Quarq-runner compatible) + frozen corpus + CI floor gate proving the memory-first stack vs 98.2% |
| AU-AHE.harness.failure-evolution | Failure-Driven Evolution | Langfuse failure ingest → failure_gap Concept topics → golden-loop remediation → regression-gated merge |
| AU-AHE.optimization.performance-anomaly-consumer | Performance Anomaly Consumer | Turns persisted PerformanceAnomaly nodes into evolution topics (adaptation/anomaly_consumer.py) |
| AU-AHE.harness.promotion-governance-validator | Promotion Governance Validator | Governed validation gate every promoted proposal must pass (research/promotion_governance.py) |
| AHE-3.21 | Evolution-to-Branch Bridge | Change synthesis + RLM-sandbox validation + ActionPolicy-gated ChangePublisher publishing promoted proposals as reviewable local git branches |
| AU-AHE.harness.per-task-adaptation-speed | Adaptation-Speed Metric | SAI primary measure — per-task time-to-target + sample-complexity + learning-AUC over a verified-reward curve (harness/adaptation_speed.py) |
| AU-AHE.harness.sai-task | Specialization Task + Verifier Contract | The (task, verifier, target, human-baseline) contract + machine-verifiable Verifier protocol every specialization track shares (harness/sai_task.py) |
| AU-AHE.harness.sai-controller | SAI Factory Controller | Closed scaffolding+weights specialization loop steered by adaptation speed, ratchet-gated promotion (research/sai_factory.py) |
| AHE-3.28 | GPU/Compute-Kernel SAI Specialization Verifier | The first concrete machine-verifiable specialization domain — runs untrusted candidate compute-kernel source in a sandboxed separate process against a correctness+speed verifier (data-science-mcp/kernels/) |
| AU-AHE.reward.search-task-corpus | Realized Search-Difficulty Signatures | Trajectory signatures — solving cost, answer hit time, prior-shortcut rate — that diagnose whether a deep-search task forced real search and gate task acceptance (graph/training_signals.py; distills FORT-Searcher arXiv:2606.12087) |
| AU-AHE.reward.cache-rollout-signals | Cache-Tier-Aware Reward Shaping | Reward half of CacheRL — token_cache_mask masks injected tool observations from the loss (only model thoughts/actions train) and cache_tier_aware_reward discounts failures caused by low-reliability fuzzy/semantic cache tiers (graph/training_signals.py; distills arXiv:2606.14179) |
| AU-AHE.harness.co-evolve-research | Hybrid Tri-Evolution Controller | Co-evolves the research proposer/solver/judge with interdependent rewards and proves co-evolution is indispensable via a joint-vs-solo ablation (harness/hote_tri_evolution.py; distills HOTE arXiv:2606.13710) |
Pillar 4: Ecosystem & Peripherals (ECO-4.0 – 4.34)¶
| ID | Concept | Description |
|---|---|---|
| ECO-4.0 | Tool Interface & MCP Factory | MCP server factory, skill loading, tool assignment |
| ECO-4.1 | A2A Network & Consensus 🔬 | Agent-to-agent discovery, delegation, consensus |
| ECO-4.6 | Web/Social Content KG Ingestion | Ingest external web/social content (posts, articles) into the KG as documents for assimilation |
| ECO-4.7 | Infrastructure Telemetry Monitoring | Connector-sourced infrastructure/service telemetry ingestion and monitoring into the KG |
| AU-ECO.toolkit.journey-map-narrative | Community Telemetry & Ecosystem Map | Ecosystem topology, 40-repo graph, telemetry |
| AU-ECO.ui.company-infrastructure-orchestration | Market Data KG Node Models | Connector/fetch-record ontology nodes for data-source provenance |
| AU-ECO.toolkit.journey-map-adoption | KG MCP Server & Execution | KG MCP exposure, durable execution, sandbox |
| AU-ECO.mcp.toolkit-live-discovery | Dynamic Capability Ingestion & Discovery | Ingests external agent toolkits, discovers MCP endpoints in real-time, and builds self-documenting skill-graphs |
| AU-OS.deployment.infra-orchestration | Domain Workflow Bindings | Parallel execution workflows and capability bindings for specialized domain processes |
| AU-ECO.bus.pluggable-queue-backend | Queue Backend | Abstract QueueBackend with Memory, Nats, and Kafka implementations for multi-scale event distribution |
| AU-KG.memory.team-startup-context | Automated Documentation & AGENTS.md Governance | Deterministic hierarchical AGENTS.md management, self-improving reflectors, and codebase map generation |
| AU-OS.governance.lint-enforcement-hook | Deterministic Lint Enforcement Hook | PRE_TOOL_USE subprocess hook for ruff/mypy/eslint enforcement |
| AU-ECO.toolkit.self-documenting-plugin-bundle | Plugin Bundle Distribution System | Manifest-based skill/hook/config packaging with KG registry |
| AU-OS.governance.permission-policy | Ecosystem Governance & Policy Engine | Unified engine managing permission policies, configuration staleness auditing, and governance workflows |
| ECO-4.14 | Infrastructure Blueprint Library | Library of modular, declarative system infrastructure configurations |
| AU-ECO.mcp.profile-differences-from-client | Fleet-Scale MCP Multiplexer Hardening | Per-child concurrency limits, session pools, restart-on-crash, circuit breakers, multiplexer_status tool (mcp/child_resilience.py) |
| AU-ECO.ui.fleet-http-client-library | Fleet HTTP Client Library | Shared HTTP client + auth surface that every agents/* API client adopts so transport/auth/retry live in one place (http/__init__.py, http/auth.py, http/client.py) |
| ECO-4.36 | Dynamic MCP Tool Gateway | Boots with meta-tools (find_tools/load_tools/unload_tools), KG-discovers and lazily mounts child tools at runtime via FastMCP tools/list_changed with catalog-aware, collision-free prefix assignment — solves whole-fleet tool overload (mcp/multiplexer.py) |
| AU-ECO.interop.gateway-client-sdk | Surface Gateway Client SDK | The single shared gateway client every frontend (webui, terminal-ui, geniusbot) uses, so transport/auth/retry live in one place (gateway_client/) |
| AU-ECO.connector.agent-source-ingestion | Usage Session Ingestion | Auto-detecting parser registry that turns agent session logs into normalized usage records (agentsview assimilation) |
| AU-OS.observability.usage-analytics-store | Usage Analytics Store | Backend-abstracted pydantic usage/cost models — the single source of token cost for the whole stack |
| ECO-4.40 | LiteLLM Pricing Source | Per-model token pricing feeding cost across the stack |
| AU-ECO.mcp.usage-cost-observability-surface | Cross-UI Usage & Cost Surface | The usage/cost/observability surface (/api/observability, usage_query) rendered in all three frontends |
| AU-ECO.mcp.client-side-chat-session | Remote Usage Ingest | Clients parse local logs and POST normalized usage to the server sink |
| AU-ECO.connector.git-task-resolver | Document → KG Fact Extraction UI | Interactive document/URL → atomic-triple extraction with live force-graph, edge-fact cards, and JSONL across all three frontends over /api/enhanced/extract/* |
| ECO-4.82 | Verbose 1:1 MCP Tool Surface | MCP_TOOL_MODE (condensed default / verbose / both) over a shared register_tool_surface that auto-discovers register_<domain>_tools and exposes a typed 1:1 verbose tier |
Pillar 5: Agent OS Infrastructure (OS-5.0 – 5.29)¶
| ID | Concept | Description |
|---|---|---|
| OS-5.0 | Agent OS Kernel & XDG Paths | Kernel lifecycle, XDG path resolution |
| OS-5.1 | Security & Auth | JWT/API auth, session concurrency, injection scanner |
| OS-5.2 | Resource Scheduling 🔬 | Cognitive scheduler, token quotas, preemption |
| AU-OS.governance.reactive-multi-axis-budget | OS Guardrails & Safety Boundaries | Holistic boundary definition integrating tool guards, reactive budgets/homeostasis, and ontological guardrails |
| AU-OS.governance.wasm-micro-agent-sandbox | Telemetry & Observability | OTEL, token tracking, audit logging |
| OS-5.5 | Massive Scale | Pluggable distributed queues, epistemic-graph Rust UDS RPC, and wasmtime sandbox integration |
| OS-5.14 | Server-Minted JWT Identity | ActorContext minted server-side from validated JWTs; fail-closed permissioning; HMAC engine auth (security/request_identity.py, security/auth.py) |
| AU-OS.config.fleet-event-ingress | Fleet Event Ingress | POST /api/fleet/events webhook persisting monitoring alerts as FleetEvent KG nodes + triage seam |
| AU-OS.state.unified-durable-state-externalization | Unified Durable-State Externalization | One STATE_DB_URI flag moves checkpoints, sessions/goals, and the task queue onto shared Postgres |
| AU-OS.state.cross-host-daemon-leadership | Cross-Host Daemon Leadership | Postgres advisory-lock election so singleton background ticks run on exactly one host fleet-wide |
| AU-OS.state.fleet-supervisory-plane-at | Fleet Supervisory Plane at Scale | SQL aggregation, paginated/filtered session queries, desired-state pause/kill reconciliation across hosts |
| AU-OS.observability.no-op-without-metrics | Gateway Middle-Tier Hardening | Prometheus /metrics, per-tenant token-bucket rate limiting, engine circuit breaker, GATEWAY_WORKERS pre-fork |
| OS-5.24 | ActionPolicy Decision Point | Per-action autonomy tiers, durable rate limits, maintenance windows, blast-radius caps; fail-closed; audit-logged |
| AU-OS.config.desired-state-fleet-reconciler | Desired-State Fleet Reconciler | Leader-only tick diffing the fleet registry against a FleetObserver, converging through ActionPolicy (dry-run default) |
| AU-OS.host.remediation-playbooks | Remediation Playbooks | service_down/service_flapping/resource_pressure playbooks with stepwise verification on the fleet-event triage seam |
| AU-OS.config.health-gated-deploy-rollback | Health-Gated Deploy Watch | Durable post-deploy health watch; failure invokes policy-gated rollback + escalation |
| AU-OS.scaling.shard-topology-visibility-per | Shard Topology Visibility | Per-shard reachability/breaker status surfaces + per-endpoint engine gauges and counters |
| OS-5.29 | Reactive Replica Autoscaling | Registry-declared scaling bounds + pluggable signal providers + leader-only target-tracking autoscaler behind the policy gate |
| AU-OS.identity.so-jwt-protected-children | Multiplexer Outbound Service Token | The MCP multiplexer mints + refreshes a Keycloak client-credentials token (audience agent-services) and attaches it to every remote child so jwt-enforced connectors are reachable through the aggregator (mcp/client_credentials.py, opt-in MCP_CLIENT_AUTH=oidc-client-credentials) |
| OS-5.52 | Modular Skill/Prompt Contribution | Any agent-package ships its own skills + system prompts in its wheel, discovered via the agent_utilities.skill_providers / prompt_providers entry-points (package-owned skills install into the shared XDG dir) |
| OS-5.72 | Env-var Drift Guard | Code is the single source of truth for env vars; check_env_var_drift flags DEAD / UNDOCUMENTED / MISSING_TOOL_MODE across .env.example, mcp_config*.json, docker compose, and the README (mcp/check_env_var_drift.py) |
Gateway Service Dashboard (AU-OS.config.gateway-service-dashboard)¶
| ID | Concept | Description |
|---|---|---|
| AU-OS.config.gateway-service-dashboard | Gateway Service Dashboard | Unified 50-widget dashboard data layer with registry, aggregator, REST+WS API, and MCP auto-discovery. Synthesized from former service-dashboard-core into agent_utilities/gateway/. |
Agent OS Architecture¶
The Agent OS is a multi-subsystem architecture where the Active Knowledge Graph (KG-2.0) drives all tool discovery and routing:
| Subsystem | Package | Role |
|---|---|---|
| 🧠 Kernel | agent-utilities |
Models, logic, graph orchestration, KG, default catalog |
| 🖥️ Desktop Cockpit | geniusbot |
Premium multi-platform PySide6 Systems & Finance Cockpit GUI (CONCEPT:AU-GBOT.cockpit.through-gbot) |
| ⚙️ OS Layer | systems-manager |
Host OS operations + Agent OS MCP wrappers (23+ tools) |
| 📦 Container Runtime | container-manager-mcp |
Docker/Podman lifecycle (60+ tools) |
| 🌐 Network Stack | tunnel-manager |
SSH tunnels, remote exec, file transfer (43 tools) |
| 📂 Workspace | repository-manager |
Git workspace mgmt, dependency graphs (24 tools) |
Query Lifecycle Walkthrough¶
- Protocol Ingress (ECO-4.0): Query arrives via
/acp,/ag-ui, or/a2a. - Usage Guard (OS-5.1): Validates rate limits, execution budgets (ORCH-1.3).
- TeamConfig Check (AHE-3.3): Router checks KG for proven specialist coalition.
- Planner (ORCH-1.1): HTN goal decomposition and LATS fallback logic.
- Memory Injection (KG-2.1): Fetches Virtual Context Blocks and rationales.
- Dispatcher (ORCH-1.0): Spawns Specialist Superstates in parallel.
- Execution (ECO-4.0): Specialists interact with MCP servers or Universal Skills.
- Verification (AHE-3.1): Quality scoring with feedback loop on
< 0.7. - Persistence (KG-2.0): Traces/evaluations stored into the Knowledge Graph.
Evolution Pipeline — Super-Assimilation Architecture¶
The evolution pipeline (agent-utilities-self-evolution) provides autonomous, KG-driven
assimilation of external codebases and research papers into the agent-utilities core.
Assimilation Heuristic¶
All assimilation follows the Wire or Discard principle:
- Wire-First: Every feature MUST connect to an existing hot path (≤3 hops from entry point)
- Extend, Don't Duplicate: Overlap ≥ 0.7 similarity → extend existing CONCEPT:ID
- No Dead Code: No live call path → rejected
- Constitution Preservation: External codebases' governance rules are ingested as PolicyNodes
4-Phase Pipeline¶
flowchart TD
A[Phase 1: Ecosystem Ingestion] --> B[Phase 2: Assimilation Codification]
B --> C[Phase 3: Parallel Comparative Analysis]
C --> D[Phase 4: SDD Plan Generation]
subgraph P1 [Phase 1: Ingestion]
A1[AU-ECO.mcp.toolkit-live-discovery: agent-packages] --> A3[ORCH-1.0: IntelligencePipeline]
A2[AU-ECO.mcp.toolkit-live-discovery: open-source-libraries] --> A3
A3 --> A4[KG-2.2: PolicyIngestor: Constitution Rules]
end
subgraph P3 [Phase 3: Analysis]
C1[ORCH Background Research] --> C6[Synthesis]
C2[KG Background Research] --> C6
C3[AHE Background Research] --> C6
C4[ECO Background Research] --> C6
C5[OS Background Research] --> C6
C6 --> C7[KG-2.2: Concept Cross-Reference Matrix]
end
subgraph P4 [Phase 4: SDD]
D1[Feature Recommendations] --> D2[Wiring Audit]
D2 --> D3[KG-2.2: Constitution Compliance]
D3 --> D4[AU-ORCH.planning.legal-automation-roadmap: SDD Implementation Plan]
endIntegration Points¶
| Component | Role in Evolution |
|---|---|
PolicyIngestor (KG-2.2) |
Ingests external constitutions as PolicyNodes |
IntelligencePipeline (KG-2.0) |
Bulk codebase ingestion via graph-os MCP native ingestion |
graph_analyze (KG-2.0) |
Parallelized engine + OWL analysis per pillar |
concept_map.md |
Source of truth for 70 canonical concepts to cross-reference |
constitution.md |
Assimilation Governance rules enforced during SDD |
KG Node Types¶
| Node Type | Purpose |
|---|---|
EvolutionCycle |
Tracks each evolution pipeline run with metrics |
SDDPlan |
Generated implementation plan from analysis |
ResearchTopic |
Topics detected for research scanning |
PolicyNode |
Constitution rules from ingested codebases |