Agentic AI Orchestration Framework for Spatial Intelligence: Supply Chain Deployment

1 Introduction

• Disconnected GIS platforms, feeds, sensors, devices, databases, and legacy systems force analysts to reconcile data instead of focusing on spatial reasoning and decision support.Agentic AI can delegate coordination tasks through planning, memory, and tool orchestration [16] [18].Source: §1 · p. 1
• Prior approaches use retrieval-augmented generation or explicit workflow graphs, but agents can degrade on multi-step tasks requiring precise parameter inference and tool sequencing.The cited work addresses topology, proximity, spatial dependencies, or workflow dependencies; [1] [2]; [8] [19].Source: §1 · p. 1
• ANASSA integrates spatial reasoning, execution-aware orchestration, feedback-driven adaptation, and hallucination governance in an autonomous geospatial-intelligence layer.The paper describes the architecture as instantiated in an operating supply-chain deployment.Source: §1 · p. 1
• The paper contributes an eleven-component, four-layer architecture with typed contracts, a supply-chain deployment, and protocols for controlled mechanism-level evaluation.The deployment assesses multi-hazard freight-lane risk, including cold-chain excursions from container telemetry; the paper also identifies design principles for agentic GIS.Source: §1 · p. 1
• The deployment outcomes are observations against a pre-existing manual workflow, not controlled experiments, and do not identify which architectural mechanism produced improvement.The paper points to threats-to-validity analysis, instrumentation, and ablations as steps toward attributable evidence.Source: §1 · pp. 1–2

2 ANASSA Framework

• Agentic GIS extends earlier GIS workflows with reasoning, multi-agent execution, feedback, and validation, while the field's approaches remain fragmented in spatial reasoning, execution reliability, and evaluation.Multi-agent decomposition remains sensitive to ambiguous constraints and long dependency chains, while structured reasoning does not fully integrate execution feedback [1] [2] [19].Source: §2 · p. 2

2.1 Architecture Components

• ANASSA organizes eleven components across four layers, with C8 providing trusted authoritative data and C9 and C11 closing downward and upward process flows.C5 is the Cognitive Engine; C4-L0 through C4-L4 are infrastructure levels internal to C4, not additional architectural layers, and A0–A4 are a separate coordination-level scheme.Source: §2.1 · p. 2

2.2 Geospatial AI Cognitive Loop

• C5 is a six-step feedback cycle from sensing through learning, coordinated by C4-L0 and grounded in authoritative C8 data.Execution outputs feed back into learning; authorized operational parameters may adapt, while governance-policy changes require explicit approval [19].Source: §2.2 · p. 2
• Blocking human approvals occur at planning and before publication of cross-validated outputs, with approver identity and timestamp recorded.The gates are intended to keep consequential geospatial decisions under explicit human accountability, at a cost to machine-pace orchestration.Source: §2.2 · pp. 2–3

2.3 Governance and Hallucination Management

• C7 overlays the framework with seven governance pillars and has a Claim Validation Agent test spatial claims against certified C8 ground truth.Claims that do not match certified assertions accompany the output on an explicit unverified-claim list; this mechanism is specified as contract K5.Source: §2.3 · p. 3

3 Research Questions and Use Case

• The use case combines existing telemetry, network, and hazard streams for multi-hazard supply-chain risk assessment, without exercising cross-domain generality.It requires no new sensing infrastructure and examines the orchestration and spatial-reasoning layer on top of those streams.Source: §3 · pp. 3–4, §3.1 · p. 4
• The four research questions examine coordination effort, spatial-claim grounding, auditable multi-source fusion, and whether blocking human approvals preserve accountability without eliminating latency advantages.The questions are scoped to a single instantiated supply-chain risk-assessment use case.Source: §3 · pp. 3–4

3.1 Use-Case Requirements and Failure Model

• Typed contracts expose missing layers, coordinate reference systems, unresolved owners, expired restrictions, ungrounded facilities, and trajectory gaps at handoffs instead of allowing silent downstream degradation.The Orchestrator monitors execution and replans when an intermediate output or dependency is missing.Source: §3.1 · p. 4

3.2 Supply-Chain Agent Specification

• The coordination specification distinguishes A0 orchestration, A1 cognitive-loop roles, A2 domain specialists, A3 governance, and A4 authoritative foundation services.In this deployment, A4 is provided by C8 services rather than a separate agent.Source: §3.2 · p. 4

3.3 Coordination Semantics

• A3 governance can block or return work, while A1 and A2 cannot bypass its decisions; validation conditions make operating constraints checkable while consequential judgment remains at HITL gates.Checks include certified facility geometry, gap-aware trajectories, validity-bounded restrictions, and cited observations for reallocation proposals.Source: §3.3 · p. 4

4.1 Deployment Architecture

• Telemetry is a high-cardinality consignment-keyed point stream joined spatio-temporally with gap-aware tracks and asynchronously updated hazard polygons using version-recorded spatial indexes.Typed contracts declare the join once rather than requiring analysts to reconstruct it for each query.Source: §4.1 · p. 4

4.2 Typed Contracts

• Every agent boundary uses a typed contract with declared preconditions, and machine checks validate fields, types, dependencies, and CRS/temporal requirements while operational appropriateness remains a human judgment.Structured spatial claims are checked against certified evidence; unmatched claims remain explicitly unverified.Source: §4.2 · pp. 4–5

4.3 End-to-End Execution Trace

• The cold-chain trace versions and certifies incoming telemetry before packaging CRS and monitoring gaps, then routes a planned task graph through a blocking approval before execution.A second blocking approval precedes publication; approved outputs retain query and spatial-index version provenance, and learning updates enumerate changed parameters.Source: §4.3 · pp. 5–6

4.4 Multi-Agent Coordination of a Consignment

• For one consignment, the domain roles cover origin, port states, maritime corridor, inland routing, and inventory allocation; an expired restriction blocks execution and returns the workflow for replanning.The corridor geometry is the same versioned artifact later tested against certified terminal geometry under K5.Source: §4.4 · p. 6

4.5 Cold-Chain Excursion Detection

• For the illustrated canned-food POC, 95°F was the configured maximum and a selected 93°F reading was surfaced as an early warning before that threshold.Sampling cadence is source-determined; missing windows are treated as monitoring gaps, and event extent follows valid trajectory segments rather than inferred paths across missing measurements.Source: §4.5 · p. 6

4.6 Versioning, Provenance, and Reproducibility

• Versioned telemetry, hazard layers, and indexes tie the artifact used by a domain agent to the state later checked by Claim Validation, but audit re-issue does not reproduce the full agentic reasoning process bit-for-bit.Quantitative re-query determinism was not recorded, so RQ3 remains illustrated rather than experimentally supported.Source: §4.6 · pp. 6–7

5.2 Threats to Validity

• The outcomes are uncontrolled observations from a three-month POC with three implementation iterations, not independent experimental runs or inferential estimates.The POC was not randomized, lane conditions varied, and seasonal, volume, and iteration effects cannot be separated; deployment evidence covers one operating context and lacks intervention-rate and gate-dwell-time instrumentation.Source: §5.1 · p. 8, §5.2 · p. 8

5.1 Observed Outcomes

• 25% lower unsupported-claim rate against the manual workflow was associated with the POC.Data quality is defined as unsupported published claims divided by total published claims; unverified claims routed to review are tracked separately.Source: §5.1 · p. 8
• 50% reduction in workflow runtime relative to the preceding manual workflow was associated with the POC.Runtime spans initiation of the multi-hazard workflow through production of the validated output.Source: §5.1 · p. 8
• 60% lower decision latency was associated with the deployment, measured from objective to approved decision.The interval includes both blocking approvals, so it measures the human-governed end-to-end path.Source: §5.1 · p. 8

5.3 Instrumentation for Controlled Evaluation

• Contract, gate, and provenance logs can support evaluation of analyst effort, spatial errors, fusion coverage, audit completeness, interventions, gate dwell, and decision latency.Unverified-claim list length is also important because low authoritative coverage can shift coordination burden to human review.Source: §5.3 · p. 8

5.4 Industrial Impact

• The architecture is aimed at organizations whose analysts spend substantial time reconciling systems, but its transfer beyond the single instantiated domain is an applicability claim, not a demonstrated result.Transfer depends on adequate authoritative coverage that keeps reviewable claims at operational volume and on domain experts declaring typed boundaries up front.Source: §5.4 · p. 8

6 Discussion

• RQ1 and RQ2 receive partial support, while RQ3 lacks quantitative fusion and determinism measures and RQ4 lacks intervention-rate and gate-dwell measurements.Analyst time and systems touched were not separately recorded, grounding's contribution was not isolated, and blocking gates did not eliminate the observed latency advantage.Source: §6 · pp. 8–9

6.1 Design Principles Operationalized

• Reasoning and execution are coupled in C4-L0, while K1–K8 make declared transition preconditions checkable across domain boundaries.What transfers is the typed boundary rather than the logistics model.Source: §6.1 · p. 9
• C5-S6 and K8 require explicit cycle identifiers and changed-parameter enumeration, while C8 and K5 route unmatched claims to review before publication.These principles pair execution-time correction and explicit learning updates with authoritative grounding before publication.Source: §6.1 · p. 9
• Analyst time, spatial-reference errors, provenance completeness, intervention rate, and gate dwell are paired with principles as measurements, not asserted causal outcomes.Aggregate deployment observations cannot isolate individual mechanisms.Source: §6.1 · p. 9

6.2 Limits and Boundary Conditions

• Validation conditions externalize some expert judgment but do not capture tacit competence; safely raising autonomy requires systematically eliciting and representing organizational decision knowledge.HITL gates preserve that expertise in the operating loop rather than claiming to encode it.Source: §6.2 · p. 9
• Effective GIS-tool integration remains an engineering assumption, and the architecture does not itself solve interoperability, robust error detection, or production scalability.No general cost or throughput envelope is claimed; utilization and commercial cost figures are confidential.Source: §6.2 · p. 9
• If authoritative coverage cannot resolve enough claims at operator-accepted precision, the unverified-claim list can make human review a bottleneck and relocate coordination burden.Evidence from one operating supply-chain context does not establish cross-domain generality.Source: §6.2 · p. 9

• A grounding ablation would disable K5 and Claim Validation while retaining contracts and gates, measuring spatial-reference errors, unsupported claims, review load, and failure-detection stage.Source: §7 · p. 9
• A governance ablation would make K4/K6 advisory, while a contract ablation would replace typed contracts with ad hoc handoffs and retain the other mechanisms.The proposed measures include claim errors, interventions, gate dwell, latency, analyst time, systems touched, and contract failures.Source: §7 · p. 9
• Further evaluation should use GeoAgentBench [19] and GeoBenchX [8], vary data-certification regimes, and test gate and unverified-claim interfaces with logistics planners.Source: §7 · p. 9
• Broader research should study checkable representations of tacit decision knowledge, multi-step spatial reasoning, and modalities including remote sensing and temporal data.Source: §7 · p. 9

8 Conclusion

• In one operating supply-chain deployment, ANASSA was associated with 25% improved data quality, 50% lower spatial workflow runtime, and 60% lower decision latency against a manual baseline; observations were uncontrolled.Source: §8 · p. 9
• The deployment outcomes are uncontrolled observations and do not establish which architectural mechanism produced the improvements.Instrumentation and ablations are specified as a path toward controlled evaluation.Source: §8 · p. 10
• ANASSA automates coordination labor rather than replacing spatial expertise, retaining accountable human decisions where consequences are highest.Source: §8 · p. 10