Reliable Agentic Workflow Architectures

Abstract

An agentic system becomes useful when it turns an open-ended request into controllable operations and verifiable results. This note examines the building blocks described in ReAct, AutoGen, SWE-agent and Agentless, then proposes an architecture for documentary and spatial tasks. The contribution is methodological: define task contracts, separate evidence from interpretation, bound loops and evaluate results against a simpler procedure. No deployment or performance improvement is claimed. The reliability sought concerns a precise conclusion and its supporting evidence, rather than the fluency of the report alone.

The problem: a plausible chain can still be wrong

A convincing answer can assemble accurate sources around an incorrect conclusion. In a site study, one document addresses an earlier phase, a map shows nearby infrastructure and a schedule concerns an administrative procedure. The final text may present them as evidence that a grid connection is available. Each element exists; their combination does not justify the decision.

The difficulty increases when successive steps pass summaries to one another. A caveat disappears, a unit changes, and a subsequent agent treats that output as an observation. The system must therefore preserve verifiable records between steps: source, excerpt, scope, value, unit and status. Free text can accompany a record without replacing its fields.

ReAct, Yao et al. (2022) combines reasoning and actions with feedback from the environment. This interaction provides a mechanism for revising the plan; it does not guarantee that every observation is interpreted correctly. In this proposal, a tool call is an acquisition step followed by explicit validation.

Proposed architecture: bounded autonomy

The starting point would be a task contract. It specifies the question, expected outputs, permitted sources, execution budget and stopping criteria. “Study this site” remains too vague. “Identify the documents needed to test a grid-connection hypothesis within a defined boundary” makes it possible to check coverage without promising available capacity.

An orchestrator would maintain shared state and delegate bounded tasks. A collector would find documents; an extractor would produce observations; an analyst would construct relationships; a reviewer would look for contradictions. These roles are logical responsibilities: they do not necessarily require four models or four conversations.

AutoGen, Wu et al. (2023) provides a conversation framework involving customizable agents, tools and human participants. This coordination capability motivates the study of specialized roles. Our additional design choice would be to route results through a structured register, so that conversation does not become the system’s only memory.

The register would distinguish three statuses. An observation reproduces what a source actually establishes. An inference states a deduced relationship, its assumptions and its dependencies. A confirmation records the response of the competent authority or professional to a specific question. An agent could not automatically convert an inference into a confirmation.

Transitions would be controlled by rules independent of the generated text. A numerical observation without a unit would remain incomplete. Geometry without a coordinate reference system would not be used for distance measurements. A conclusion based on incompatible versions would return to investigation. These rules would reduce detectable ambiguity without claiming to cover every semantic error.

SWE-agent, Yang et al. (2024) examines the importance of interfaces designed for model actions in a software environment. The useful analogy is the tool contract: a territorial search should return a scope, coverage status and explicit errors, rather than a simple statement that the search succeeded.

Choose the workflow before the number of agents

A fixed sequence would suffice for a known document: extraction, field validation, comparison and drafting. A more open loop would be useful when the next source depends on what has just been found. The degree of autonomy should respond to that uncertainty rather than a preference for an impressive architecture.

Agentless, Xia et al. (2024) explores software localization and repair without allowing the model to choose future actions freely. This work encourages retaining a simple baseline in any comparison. It does not demonstrate that an identical procedure would solve a physical-infrastructure study.

Our proposal distinguishes predefined paths from dynamic control. The design rule to test would be to start with the least complex path capable of meeting the contract, then add a loop only to address an identified error or gap.

Failure modes to make visible

The first failure would be artificial convergence: several agents repeat the same error because they use the same excerpt or assumption. The number of opinions then does not measure the independence of the evidence. The reviewer should seek documentary contradictions and disclose shared sources rather than vote for the majority formulation.

The second would be an unproductive loop. A retry reformulates the question without acquiring new information. The orchestrator should track new acquisitions, stop repetitions and list the remaining unknowns. The third would be confusing absence with unavailability: an inaccessible source does not prove that a document or infrastructure does not exist.

A fourth risk concerns instructions embedded in the documents being consulted. Their content must remain data to analyze. An instruction found in a PDF should change neither permissions nor the task. Finally, an incorrect association between two parcels, companies or phases can survive every syntax check; it requires verification of identifiers and context.

Proposed evaluation protocol

Evaluation would begin with a frozen corpus of public dossiers and an assessment rubric annotated by competent readers. Each dossier would contain answerable questions, contradictions and genuinely unresolved unknowns. The rubric’s authors would separate certain observations from points where several interpretations remain defensible.

Three variants would be compared: extraction with rules, a sequential workflow and an architecture with delegation. They would receive the same sources and the same question. Outputs would be assessed blind for claim accuracy, coverage of relevant documents, preservation of caveats and the rate of incorrectly attributed confirmations.

Cost, duration and human interventions would be recorded alongside failures. Source-removal, unit-change and document-revision tests would probe the system’s sensitivity. An improvement would be accepted only if it survives repeated runs and does not worsen the most consequential errors. This protocol has yet to be implemented.

Limitations

This architecture replaces neither an electrical study nor regulatory validation. It organizes the preparation of a decision and makes its dependencies reviewable. The cited work mainly concerns textual or software environments; transferring it to land, energy and spatial constraints is a research hypothesis. The first evidence to produce would be a dossier that is more accurate and easier to review than one prepared through a simple procedure.