Lorikeet AI, CSS contrast-color, Codex /goal

Lorikeet uses a dual‑agent system, a Concierge that handles tickets and a Coach that configures and monitors the AI, to deliver customer‑support bots that know when to defer to humans. The team emphasizes ‘AI humility’ and domain‑specific guardrails, enabling integration with platforms like Zendesk while keeping agents safe in regulated sectors.

The new CSS contrast‑color() function computes either black or white text color based on a given background, eliminating the need for JavaScript libraries and build‑time calculations. Running during style computation, it instantly adapts to theme changes, helping reduce the 70% of sites that still fail basic WCAG contrast checks.

Eric Seufert discusses how building generative AI models, especially Meta's foundational models, shapes the future of AI, and argues that a deep grasp of advertising economics can drive a more optimistic outlook for humanity.

Rig generates a local‑first semantic graph of a project and serves it as an interactive map that AI coding agents can query. Running a single npx command indexes files, symbols, and dependencies, enabling agents to answer impact or related‑code questions without blind grepping, all on the developer’s machine.

Grove provides a self‑hosted MCP server that makes a git‑backed Obsidian vault searchable and writable via any MCP‑compatible AI client such as Claude or ChatGPT. It offers BM25‑plus‑vector search, note read/write APIs, and versioned markdown files, letting product teams integrate AI agents with their knowledge base safely.

Lenny walks through Codex’s new /goal command, showing how it turns AI into an autonomous agent that can run for hours on complex tasks. He demonstrates three real use cases, removing Sentry errors, cleaning 3,900 emails, and organizing Linear tickets, while teaching a six‑part framework for writing effective, measurable goals. The episode also covers limits and best practices.

Parastore is an open‑source retail simulation where AI‑driven synthetic consumers roam a 3D virtual store, make purchase decisions, and follow behavioral patterns. It enables product placement, layout, and acquisition testing, achieving a 0.955 Spearman correlation against real POS data. The stack uses Python/FastAPI backend and React/Three.js frontend.