AI Architecture Updates: September 15, 2026

1. Postgres Can Carry Durable Workflow State Without Temporal or Step Functions

InfoQ. Raman Varma lays out a pattern that makes Postgres the coordination layer for long-running workflows instead of adding a stateful external orchestrator. SELECT ... FOR UPDATE SKIP LOCKED turns a table into a work queue where each execution is claimed by exactly one worker, giving exactly-once processing without a message broker or distributed lock. Step checkpoints use an (execution_id, step_id) primary key with ON CONFLICT ... DO NOTHING, so a recovering worker reads the prior result rather than re-running the step. A lease-and-sweeper pattern handles crashes, and sleeps and human approvals persist as rows with wake_at timestamps rather than in-memory timers. Source

2. Alex Porcelli Argues High-Stakes Agent Decisions Belong in DMN Models, Not Prompts

InfoQ. Presenting at QCon AI Boston, the Aletyx CEO frames non-deterministic output and missing accountability as the gap blocking enterprise agent deployment, and proposes moving decision logic into DMN, an open standard for executable business rules. The pattern has three layers: business users define decision tables with AI help converting spreadsheets into executable models, the DMN engine generates skill definitions and REST endpoints automatically, and agents invoke those models as deterministic tools rather than generating ad-hoc answers. He demonstrates NVIDIA NeMo guardrails intercepting LLM output and routing high-stakes decisions through DMN execution first, and reports the engine processing 18,000 transactions per second. Source