Databricks Links Temporal and Lakebase for AI Agents

Databricks has introduced a new architectural blueprint that integrates Temporal's durable execution engine with Lakebase Postgres to solve the fragility of long-running AI agents. In cloud environments, agentic workflows can easily fail due to routine worker restarts, container redeployments, or API timeouts. By combining these two systems, developers can ensure that an agent preserves its state, retries failed operations safely, and pauses indefinitely for human intervention without consuming active compute resources.

The reference implementation demonstrates this architecture through a mock personal-loan underwriting agent. In this setup, Temporal manages the workflow's control-flow history, while Lakebase Postgres serves as the queryable operational database for application state. To handle transient errors, the system configures distinct retry policies: up to four attempts within a three-minute timeout for model calls, up to three attempts within a 60-second timeout for tool calls, and up to five attempts within a 15-second timeout for Lakebase database writes.

Data governance is maintained by syncing underwriting policies directly from Databricks Unity Catalog to Lakebase. To ensure auditability, the architecture leverages the Lakebase Change Data Feed, currently in public preview, which flushes operational updates from Postgres to Unity Catalog Delta history tables roughly every 15 seconds. Security is handled via OAuth machine-to-machine authentication, with the client automatically refreshing its SQLAlchemy connection pool before the one-hour database credentials expire.

For AI practitioners, this pattern provides a robust framework for building complex, human-in-the-loop applications. Instead of losing progress during a system crash, a replacement worker can instantly resume an agent's execution from its last recorded step. The accompanying test suite includes 21 passing tests covering workflow sequencing, OAuth connections, and idempotent persistence, offering a production-ready starting point for resilient enterprise automation.