- The paper introduces an autonomous coding paradigm that seamlessly integrates data interpretation, schema creation, and query generation while producing auditable artifacts.
- It employs a modular agent design with execution-grounded verification across multiple SQL dialects, achieving significant performance gains on key benchmarks.
- The study emphasizes reliable workflow compression and persistent memory sharing, despite increased latency from iterative generation and verification loops.
Data Intelligence Agents: An Autonomous Coding Paradigm for Enterprise Data Interpretation, Modeling, and Querying
The persistent friction in enterprise data workflows stems not from the lack of data, but from the repeated, lossy handoffs required to transform raw sources into actionable insights. Conventional pipelines fragment the process across discovery, schema construction, and query generation, each step incurring context loss and latency between domain experts, engineers, and analysts. Existing LLM-powered solutions typically address isolated fragments such as text-to-SQL, conversational querying, or schema inference, but none fully compresses the multi-stage data intelligence loop while preserving auditable, executable artifacts for expert review. DIA targets this gap, positing that autonomous coding agents (ACAs) — execution-capable, memory-aware counterparts to LLMs — should be the central abstraction for data intelligence, not merely text generation.
System Architecture
DIA operationalizes three distinct agents, all realized within a single ACA instantiated via the OpenHands framework powered by Claude Sonnet 4.5. These agents — Data Interpreter, Schema Creator, and Query Generator — interact within a shared workspace, consuming and producing executable artifacts instead of ephemeral text. Artifacts are persisted for domain expert audit, enabling both operational reliability and explainability. The system’s memory hierarchy (episodic, session, cross-session tiers) is artifact-based and pull-driven, ensuring only execution-verified knowledge informs downstream behaviors.
- Data Interpreter: Profiles heterogeneous raw sources, inferring schemas, key candidates, join paths, and quality anomalies through executed code. The result is a structured interpretation artifact consumed directly by the Schema Creator.
- Schema Creator: Materializes relational databases from interpreted profiles, enforcing key, constraint, and integrity checks with a load-first-normalize-second discipline. Artifacts include database schema manifests, validation reports, and reject buffers for integrity violations.
- Query Generator: Translates NL questions to executed SQL (across four dialects), with iterative shape declaration, schema probing, candidate generation, execution, and self-verification. Adaptation is limited to in-context natural language scaffolding; no task-specific fine-tuning is performed.
Empirical Evaluation
The Query Generator is evaluated on seven public SQL benchmarks (BIRD-Dev, BIRD-Critic, LiveSQLBench, BIRD-Interact, Spider2-Lite, Spider2-Snow, Spider2-DBT), covering four task categories (generation, debugging, conversational, dbt project completion) and four SQL dialects. Key outcomes:
- Outperformance Across Benchmarks: DIA matches or exceeds the best published results on all seven, with margins as high as +33.0 on BIRD-Interact, +16.1 on Spider2-Lite, +15.4 on BIRD-Critic, and +12.7 on LiveSQLBench. On saturated benchmarks like BIRD-Dev, it equals RL-trained specialists.
- Dialect Generalization: DIA's architecture, execution-grounded self-verification, and memory mechanisms achieve robust performance across SQLite, PostgreSQL, Snowflake, and DuckDB, outperforming both agent-based and training-based baselines.
- Structured Task Strength: Relative gains are most pronounced on modification and management tasks due to the agent's explicit shape declaration and validation, whereas high-level, composite metric queries remain challenging.
- Error Taxonomy: Semantic reasoning, output convention inconsistencies, grounding (identifier misalignment), and rare execution failures constitute dominant error modes. Execution grounding eliminates most syntactic failures; semantic gaps persist where user intent outstrips database evidence.
Memory and Experience Reuse
DIA's memory subsystem accumulates and promotes conditional rules grounded in execution evidence, spanning aggregation strategies, output conventions, join path selections, and value normalization. Promotion from session to cross-session lessons is outcome-gated, preventing test leakage and ensuring auditability. The memory store enables both immediate within-workspace recall and long-term knowledge sharing across agents, facilitating rapid adaptation to domain-specific recurrence structures.
Deployment and Production Insights
DIA is actively deployed with enterprise customers. Its end-to-end workflow enables domain experts to upload raw operational data, receive structured interpretations, validate schema manifolds, and obtain data analyses without writing SQL or DDL directly. All steps are artifact-driven, audit-ready, and allow expert intervention, catering to both operational rigor and transparent workflow compression. The production architecture scales across enterprise source systems, leveraging underlying data platform primitives for ingestion, access control, and execution.
Limitations and Forward Directions
DIA trades computation for reliability, with iterative generation-execution-verification loops leading to per-question latencies potentially unsuitable for strict interactive or high-throughput scenarios. Verification remains execution-grounded; latent semantic mismatches persist when agent misinterpretation aligns both query and check. Expansion targets include:
- Wider Benchmarking of the Data Interpreter and Schema Creator on data preparation tasks;
- Model Diversity: Sensitivity analysis across different LLMs;
- User-Centered Validation: Incorporating real user interactions over simulated protocols;
- Memory Structuring: Migration from file-based artifact stores to graph-structured knowledge representation for more effective mining and retrieval.
Implications and Future Outlook
DIA establishes the ACA as the foundational abstraction for enterprise data intelligence, supplanting brittle text-only LLM workflows. Its artifact-driven, execution-grounded, memory-aware architecture proves generalizable across diverse benchmarks and enterprise tasks with minimal adaptation. The theoretical implication is a shift towards code-centric agents with persistent, auditable memory for all stages of the data lifecycle, rather than task-specialized, text-generating models. Practically, DIA reduces latency, context loss, and operational risk in enterprise analytics, facilitating reliable, explainable, and auditable data workflows. Future AI systems are poised to further abstract away the boundaries between data discovery, schema modeling, and querying, consolidating them into unified memory-augmented autonomous coding frameworks.
Conclusion
DIA demonstrates an ACA-centric systems design for enterprise data intelligence, validated across heterogeneous tasks, dialects, and benchmarks. By grounding workflow stages in execution, artifact persistence, and shared experience, DIA achieves benchmark-leading reliability and generalization with a single LLM and no fine-tuning. The architecture's practical efficacy and theoretical coherence signal a paradigm shift in AI-driven enterprise analytics, with artifact-centric coding, execution-grounded verification, and persistent memory as foundational principles for future development (2606.19319).
