Agent DB Group
Overview
Agent-centric data systems
The Agent DB Group studies how data systems can support agentic AI workflows that evolve across model calls, tools and external resources. Our Agent-Centric Data System (ACDS) reference architecture organizes these workflows into four task-processing layers—Perception, Planning, Execution and Action—supported by Memory, a Data Interface and Inference Infrastructure.
Our research focuses on the interfaces between these layers: preserving task records and provenance, adapting workload plans, managing cross-system effects and recovery, and evaluating where errors arise across an agent workflow.
Architecture layers
- Perception Layer. Interprets incoming requests using the available task context and data, producing perceived queries with references to the evidence behind them.
- Planning Layer. Organizes perceived queries into a workload, shares compatible operations, and preserves ordering and isolation when producing scheduled queries.
- Execution Layer. Assigns scheduled work to agents and specifies invocation dependencies, execution stages, and verification or recovery branches.
- Action Layer. Resolves and validates tool and data operations, carries them out, and records verified outcomes and evidence of external effects.
- Memory Layer. Manages task records across steps and agents, selects authorized context, and retains links to the data on which those records depend.
- Data Interface Layer. Exposes heterogeneous resources as source-backed DataObjects, preserving access conditions, supported operations and provenance while persisting task records.
- Inference Infrastructure Layer. Coordinates compute, storage and communication for model calls, data operations and managed state while respecting agent assignments and dependencies.