Agent-Aware Middleware for Autonomous Legacy System Enhancement: A Domain-Agnostic Approach

Enterprise compliance and operational tracking across educational institutions, information technology (IT) operations, and financial administration still rely heavily on rigid, file-locked tabular registers (Microsoft Excel .xlsx multi-sheet workbooks) and local relational databases (sqlite3) that lack modern API orchestration. While broader industry modernization efforts are motivated by historical legacy backbones—such as COBOL mainframes, MUMPS clinical globals, and SCADA telemetry historians—even contemporary file-backed and relational enterprise registries suffer from a severe operational divide. Traditional Robotic Process Automation (RPA) scripts rely on brittle rule paths that fracture whenever users submit unstructured natural language explanations or when worksheet schemas shift across reporting periods. Conversely, coupling a single-agent Large Language Model (LLM) directly to stateful enterprise ledgers introduces unacceptable hazards: arithmetic threshold hallucinations, malformed JSON syntax, context-poisoned cross-entity mutations, context-window saturation, and operating-system file-lock deadlocks. To resolve this architectural gap, this paper presents SwarmCore D.A.M.A. (Deterministic Agentic Middleware Architecture), an event-driven, domain-agnostic five-agent system (Concierge, Data Engineer, Sandbox Actuator, Fixer, and Compliance Auditor) governed by a centralized Shared Blackboard state pattern (core/blackboard_orchestrator.py). Rather than passing raw tabular rows into a single prompt, our LegacyBridge module (core/bridge.py) introduces Ghost-Mode Vocabulary Funneling, stripping raw row values during macro-planning passes to compress single-sheet prompt context by 99.13% (from 5,861 tokens to 51 tokens/pass) and isolating the master summary sheet during compliance auditing (789 tokens/pass, an 86.54% reduction) while delegating vectorized cross-sheet rollups to Pandas. Before any state mutation reaches disk, SandboxActuator (agents/actor.py) enforces a formal Two-Stage Deterministic Validation Airlock across four orthogonal predicates: JSON structural syntax (V_syntax), primary-key semantic target locking (V_sem, preventing valid-but-wrong-entity mutations), worksheet column existence (V_schema), and builtin-free dynamic inequality evaluation (V_math). Furthermore, all numerical thresholds, database patch literals ('L', 'Maintenance'), and HTML Telegram notification templates ([TEMPLATE 1]–[TEMPLATE 5]) are extracted dynamically at runtime via regex-guided Retrieval-Augmented Generation (RAG) from external PDF policy manuals (TransactionalAgent in agents/transactional.py), while OS-level Ghost Write timestamp spoofing (os.utime) and a 2.0 s atomic-save settle buffer prevent infinite file-watcher loops and Windows sharing violations. We empirically evaluate SwarmCore across a scaled 1,000-row multi-departmental enterprise proxy testbed (5 departments × 5 worksheets × 40 entities) alongside a 5-cluster IT server SLA registry and a corporate finance commercial loan ledger. Across an N = 400 programmatic fault-injection taxonomy (n = 100 trials per category: ARITHMETIC_THRESHOLD, MALFORMED_JSON, SCHEMA_DRIFT, and SEMANTIC_WRONG_TARGET), SwarmCore achieves a 100.0% fault detection rate (0.0% false positives) and a 100.0% task completion yield. Our four-configuration architectural ablation study proves that while pairing a single-agent LLM with our deterministic validator alone (Configuration C2) preserves 100.0% data integrity, its fault-suite task completion yield collapses to 25.0% (70.0% overall across a 1,000-operation workload) because a passive validator cannot repair broken JSON, drifted column names, or wrong target IDs. Combining the deterministic airlock with our five-agent blackboard and FixerAgent traceback recovery loop (Configuration C4) recovers 100.0% of cognitive faults while multi-provider API routing with exponential backoff raises batch audit execution yield from 60.0% to 100.0% (5/5 departmental nodes) and reduces synchronous transaction latency from the 87.92 s v1.0 baseline to p50 = 12.57 s single-pass / 25.59 s–38.17 s end-to-end on Google Gemini 3.5 Flash-Lite and p50 = 1.53 s single-pass / 5.04 s end-to-end on Groq LPU failover nodes. Publication History & Revision Note: The foundational multi-agent architecture and proxy testbed of this research were originally developed as the authors' B.Tech Final Year Project at Tagore Engineering College (Anna University) in May 2026 and initially submitted for journal peer review (Ref: EAAI-26-15695). Following constructive reviewer critique regarding empirical fault-injection denominators and architectural ablation, this October 2026 preprint represents an expanded, peer-review-revised v2.0 edition incorporating a scaled 1,000-row multi-departmental testbed, a Zero-Hardcode Policy-RAG engine, and a rigorous $N = 400$ programmatic fault-injection ablation suite.