Agent Foundations: Context, Safety, and Team Harnesses

Designing effective Genie Agents from a single prompt shows how Databricks assembles governed Unity Catalog data, documents, and permissions into testable, repeatable Genie Agents from a single prompt. That gives outcome engineers a repeatable pattern for context engineering and governed grounding, cutting setup friction for production agents (Principles 02, 06, 11).

Launch HN: OneCLI (YC S26) – OSS sandboxed agent harness for teams gives every employee a sandboxed, policy‑enforced personal agent with centralized secrets and IdP provisioning. Use this as a blueprint for team-level agent governance and auditability so individual agents scale without swallowing org controls (Principles 03, 07, 10).

New clarifying questions in Agent Runners updates Agent Runners to ask lightweight clarifying questions up front, producing sharper first deploys and faster iterations. Treat short clarification loops as a runtime primitive: they reduce wasted cycles, improve correctness, and make agent behaviors more measurable (Principles 03, 06).

Harness launches AI agents that triage and patch vulnerabilities launches agents that detect vulnerabilities, cut false positives with AI SAST, and generate developer‑approved patches to speed secure delivery. This is a concrete agent‑in‑the‑loop DevSecOps pattern for shipping secure outcomes while preserving human approval and traceability (Principles 03, 14).

Layered data architecture turns enterprise data into a system of intelligence argues that knowledge graphs and layered architectures give enterprise AI trusted context to turn scattered data into a system of intelligence. Outcome engineers should adopt layered graphs as the context engine for agents to preserve Ground Truth and enable auditable, reliable outcomes (Principles 11, 02).