From Maps to Decisions: Agents for Physical Infrastructure

Abstract

Selecting a compute site requires connecting geography, an industrial programme and external dependencies. A map can locate these elements without demonstrating that their combination is feasible. This note proposes an agentic method for assembling evidence, checking spatial correspondences and producing a decision register. The contribution is an analysis and evaluation protocol; no experiment is claimed. Documentary observations, technical inferences and external confirmations remain distinct categories. The intended output is a verifiable conditional decision, accompanied by the questions that still prevent it from being made.

The problem: proximity does not establish feasibility

On a map, a parcel near a power line looks promising. Yet distance alone says nothing about the relevant substation, usable voltage level, studied capacity, necessary works or rights of way. The reasoning must connect a defined need to a procedure and a precise technical scope.

The same difficulty applies to land. A cadastral polygon, land under control and a buildable area describe different objects. Mixing them can produce a visually satisfactory layout based on the wrong boundary. Useful spatial intelligence begins by explaining what each geometry represents and who can confirm its scope.

An agentic architecture could help coordinate these readings. It could also accelerate incorrect associations: a consultation date could become an energization date, or a nearby installation could be attached to the wrong project. The proposed objective is therefore to reduce unsupported conclusions before attempting to increase the volume of dossiers processed.

Foundations and scope of the proposal

The PROV family, presented by Groth and Moreau (2013), organizes descriptions of the entities, activities and actors involved in producing information. It provides a vocabulary for provenance; it does not certify the truth of data. We propose using it as a foundation for linking a conclusion to its documentary and geometric transformations.

ReAct, Yao et al. (2022) shows a way to alternate planning and acquisition through actions. For a territorial dossier, this logic would inspire gap-driven research: seek the document that could change the decision rather than accumulate related documents. This transfer between domains is a proposal, not a finding of the paper.

AutoGen, Wu et al. (2023) describes programmable conversations between agents, with tools and human participation. The separation of disciplines envisaged here draws on that organizational possibility. Published results on these conversations do not, however, establish their reliability for an energy-infrastructure site study.

Proposed method: a versioned spatial dossier

The first step would be to define the object under study. The dossier would record the working boundary, intended programme, phase, power requirements and scheduling assumptions. Every unconfirmed value would carry that status. An industrial scenario would thus be identifiable without being presented as the approved project or an application actually submitted.

Collection would then produce observations linked to sources. Each entry would preserve the document, its institutional author where known, its scope, revision status and a locatable passage. The system would distinguish the document date, the date of the event described and the observation date. A later change would not silently rewrite the history.

Spatial processing would follow the same discipline. Original geometry would be preserved; reprojections, clipping and intersections would be recorded as derived operations. An intersection with a constraint zone would flag an issue for examination. It would not automatically become a prohibition, because its effect depends on the applicable text and sometimes on the precise nature of the works.

Specialized tasks would examine land, energy, grid connection, environmental issues and logistics. Each would receive the same scenario identifier and working boundary. Agents would produce more than an opinion: they would identify the observations used, missing documents and conditions that could change their conclusion.

Synthesis would then seek cross-dependencies. Moving a building may change a grid-connection corridor. Reserving space for electrical equipment may reduce the area available for water management. These interactions would be represented as relationships between scenario elements, without assigning unestablished regulatory consequences.

The output would take the form of a register: question, observation, inference, validation condition and party responsible for confirmation. “A published route appears nearby” would remain an observation. “A corridor could be studied” would be an inference. “This connection is accepted under these conditions” would require a document issued by the competent party. The wording would preserve these different levels of evidence.

Coordination and tool selection

A suitable interface should return geographic results with their coordinate reference system, coverage and any truncation. It should distinguish an empty response from an access failure. SWE-agent, Yang et al. (2024) motivates attention to agents’ working interfaces in software. Here, interface quality would become an experimental variable in the protocol.

Distance calculations, intersections and unit conversions would be assigned to deterministic functions. The model would help choose the question, interpret a document and explain dependencies. It should not improvise a geometric result from the appearance of an image or a textual description.

A fixed workflow would serve as the baseline for dossiers whose steps are known. Agentless, Xia et al. (2024) offers a research example in which a structured procedure questions the need for complex autonomy. The lesson retained for our study is experimental: actually compare delegation against a simple pipeline before justifying it through apparent sophistication.

Failure modes

A particularly dangerous error would be inferring capacity. The system could convert the existence of a line, a substation or a reinforcement project into megawatts available to a user. The proposed control would prohibit that conversion without a source explicitly addressing availability and its scope. Without such a source, it would remain unknown.

Temporal mismatches would form another family of errors. An authorization may concern an initial phase while a presentation describes a broader ambition. The difference is not necessarily a contradiction. The system should ask whether the figures describe the same object before flagging an anomaly.

Finally, a dossier may be technically coherent but incomplete: an inaccessible easement, an unknown contractual restriction or a non-public grid study. An aggregate score could conceal this weakness. The proposed method would prioritize displaying decisive unknowns and unconfirmed dependencies without turning them into misleading indicators of precision.

Proposed evaluation protocol

The test corpus would consist of frozen, annotated public dossiers covering different geometries and levels of completeness. A rubric would distinguish directly supported conclusions, acceptable assumptions and claims that cannot be confirmed from the available documents. Corpus documents would remain separate from those used to design the instructions.

The same questions would be submitted to documented manual research, a fixed automated pipeline and a system with specialized agents. Evaluation would measure correct project-to-source associations, relevant contradictions, traceability of conclusions and the ability to abstain when evidence is missing. Human review time would be included in the total cost.

Controlled perturbations would introduce a boundary change, a capacity expressed in a different unit or two versions of a schedule. Reviewers would check whether dependent conclusions are reconsidered. An ablation test would remove an essential document to observe whether the system retains a conclusion that is no longer justified. No performance is announced before these tests are conducted.

Limitations and success criterion

The bibliographic sources establish building blocks for coordination and provenance; they do not validate the proposed system as a whole. Documentary analysis replaces neither a site inspection, an electrical study nor legal confirmation. The success criterion would be a dossier that enables the decision-maker to understand what is established, what is inferred and which verification would unlock the next decision.