One AI agent on every surface, with skills loaded on demand

We ran four agents: one for chat, one that built workflows, one that built apps, and one for database questions. Each had its own prompt, tools, and run lifecycle, so a fix to one often missed the others. They're now a single agent, and what changes between surfaces is the subject it's pinned to (home, a workflow, or an app), which decides its instructions, preloaded skills, and tools.

How it works

• Skills on demand. Every conversation starts with the same base tools. Authoring tools sit behind skills, loaded when needed or automatically when you open a workflow or app, and they persist for the conversation.
• Shared tools. Per-type tools merged into resource-wide ones that behave the same everywhere.
• Test runs in chat. Tests stream live inside the conversation as an expandable card.

Quality from evals

• New code and app eval suites, built from the same prompts the New Workflow and New App dialogs send, drove a round of prompt and scaffold fixes.
• Code workflows are type-checked on the server against the real runtime library before they're saved.
• A run stops with a clear message when the same tool hits the same error five times in a row.

Also shipped

• Web search falls back across three providers when one is rate-limited.
• Interactive runs are capped per member and queue visibly instead of competing for workers.