OpenAI says “chat is dead” and plans to rebuild ChatGPT as a full-blown agent app. OpenAI is rebuilding ChatGPT into an agent-driven superapp that runs autonomous agents and integrates partner apps. This flips ChatGPT from a conversational UI to an orchestration platform you must design for — think partner integrations, agent lifecycle, and new guardrails (Principle 09).

Snowflake, Databricks and the model makers: The battle for the agentic client and AI back end. Vendors are fighting over who controls the intelligent client and the data-driven AI back end that will power agentic experiences. If you build agentic workflows this shapes where you place context, lineage, and the Graph that agents rely on — plan for platform lock-in and data-ownership tradeoffs (Principle 11).

datasette-agent-edit 0.1a0. Simon Willison ships a small, reusable toolset (view, str_replace, insert) to let agents safely edit Datasette data and plugins. This gives you a composable, auditable primitive for agent-driven data changes — integrate it to keep edits legible and reversible in production (Principles 03, 06).

A Case for Simulation-Driven Resilience in Agentic Data Systems. Researchers argue simulation-driven testing reveals metastable failure modes that only show under agent workloads and prevent incidents before production. Add agent-workload simulators to your test and CI pipelines to find emergent failures early and harden your system’s immune response (Principle 14).

10 MCP servers to connect LLMs with databases. Ten Model Context Protocol servers now let LLMs query, update, and administer databases without writing SQL. That changes context engineering: agents can read and mutate authoritative state directly, so design access controls, transactional semantics, and audit trails from day one (Principles 06, 11).