Databricks' acquisition of Electric brings PGlite—a WebAssembly build of PostgreSQL—into AI agent sandboxes, syncing with a central Lakebase PostgreSQL instance. The piece argues this is a genuine architectural shift but warns against treating agent-local databases as authoritative sources of truth. It proposes separating agent state into ephemeral, durable, and authoritative categories, introduces a C.A.L.M. framework (Changeability, Assurance, Leverage, Measurability) for evaluating trust in such architectures, and an ORBIT set of principles (Outbox First, Rate State Is Shared State, Background Is the Unit of Execution, Idempotency Before Day One, Trace Everything) for reliable execution once agents act on durable data or external systems. PGlite currently supports only single-user mode, and its sync plugin is alpha with known limitations around outbound writes and conflict resolution.
Table of contents
Why agents create a different state problemOne PostgreSQL standard, different responsibilitiesSynchronization is the real systems problemApplying the C.A.L.M. Platform TestWhen agents act, ORBIT becomes relevantWhat the PostgreSQL community should examine nextQuestions this post answers
What is PGlite and how does it let PostgreSQL run inside an AI agent sandbox?
PGlite is PostgreSQL compiled to WebAssembly, packaged for JavaScript environments so it can run inside a browser, application, serverless environment, or AI agent sandbox. It operates in memory or with local persistence, supports extensions like pgvector for local vector retrieval, and currently runs only in single-user mode with one process and one connection, since WebAssembly lacks PostgreSQL's usual fork-based concurrency model. Teams evaluating embedded database options for agent workloads can follow architecture debates like this on daily.dev.
Why did Databricks acquire Electric, the company behind PGlite?
Databricks acquired Electric to extend PostgreSQL from its Lakebase lakehouse offering down to the edge, pairing PGlite running locally inside agent sandboxes with a central Lakebase PostgreSQL instance that provides shared state and control via Electric's synchronization technology. PGlite grew from one million to 13 million weekly downloads in about 12 months, reflecting fast adoption across development tools, testing frameworks, and AI applications. Anyone tracking how AI infrastructure acquisitions reshape database choices can follow this on daily.dev.
What are the limitations of PGlite's synchronization with a central PostgreSQL database?
PGlite's documented synchronization plugin is alpha-stage, with known gaps in outbound local-write synchronization and conflict resolution, meaning the architecture vision currently runs ahead of the implementation. This makes PGlite better suited as an agent-local execution store for ephemeral or checkpointed state rather than a shared system of record, since concurrent writes across many distributed agents are not yet reliably reconciled. Engineers weighing sync reliability before adopting embedded Postgres can track these developments on daily.dev.