Thoughtworks

ThoughtWorks Flags AI-Accelerated Shadow IT; Structural Fix Proposed

The ThoughtWorks Technology Radar (April 2026) called out "AI-accelerated shadow IT" under Caution. The article argues this is one visible instance of a broader cloud resource lifecycle governance problem (create, use, modify, abandon, delete) that predates AI tooling but is accelerated by it. HackerOne triage revealed S3 bucket takeover patterns where deleted resources left ghost references that attackers could exploit. The author describes a structural fix: a mechanical "specification gate" that snapshots cloud provider state and evaluates it against a catalog of safety invariants (3,000+ invariants), blocking violations across all lifecycle phases. An open-source engine (Stave) implements this approach and the author frames it as necessary to remove a two-tier system of reviewed core workflows and ungoverned shadow workflows.

SAID: Spec-driven Framework for AI-native Design

The article argues that traditional sequential design handoffs are being replaced by AI-native, continuous workflows centered on spec-driven development. It introduces the SAID framework (Specify intent, Agentic delegation, Iterative Description, Discernment & Diligence) and highlights Anthropic’s 4D AI Fluency model (Delegation, Description, Discernment, Diligence). The piece describes how Git and spec-driven practices (Agentic SDD/SDD) are turning specs into a shared workspace that guides prototypes and production, and it outlines team archetypes and lifecycle loops (Loop A: Proposition→Prototype; Loop B: Prototype→Product). The article emphasizes designers must learn to express intent in machine- and human-readable specs rather than rely solely on visual mockups.

OLAP and OLTP Lines Are Blurring

A developer article explains how recent extensions and engine architectures are narrowing the gap between OLTP (transactional) and OLAP (analytical) workloads. It describes how extensions such as pg_lake decouple storage to cloud data lakes using Apache Iceberg while offloading analytical execution to an isolated, vectorized DuckDB process to avoid impacting the operational database. The author maps end-to-end execution flow, resource safety boundaries, and scheduling differences between macro-distributed query engines and micro-morsel (embedded/vectorized) processing engines. The post links to a detailed GitHub DeepDiveDuckDB repository for a full architecture layout. The piece is a technical analysis aimed at data engineers and platform architects.