Two Tracks, One Execution Surface
Connectors and Skills give agents the ability to act; the primitives give agents the data to act on. Keeping capability and data separate stops the two from being conflated — logic stays reusable, data stays managed, and neither leaks into the other.Pre-Computed Context Over Inline Calls
Querying external systems on every turn is slow, costly, and fragile. Instead, Pipelines pre-process and persist clean data ahead of time, so agents consume structured rows and organized files rather than raw API responses mid-conversation. The heavy data work happens on a schedule; the runtime stays light.Human-in-the-Loop Is First-Class
Human review is a designed gate inside the agent, on equal footing with any other step — not an error handler tacked on at the end. It surfaces in the App as a clear “Action needed” checkpoint that arms the reviewer with the agent’s work. The professional’s judgment is part of the workflow by design.Isolation for Data, Reuse for Logic
Skills, agents, and connectors are reusable assets — built once, used everywhere. Files, tables, and pipeline data are client-specific and never cross a Workstream boundary. This split is what lets one agent safely serve many clients: identical logic, isolated data.One Platform, Two Views
The Builder and the End-User App are two lenses over the same objects. Builders assemble; end-users run and review. Neither re-implements the other, so what you build is exactly what your team runs.Key Takeaways
- Separation of concerns: capability and data are different tracks, and keeping them apart keeps both clean.
- Speed and reliability: pre-computed data beats live calls, every time.
- Control: human review is built in, not bolted on.
- Safety at scale: shared logic, isolated client data.
- Consistency: one set of objects, surfaced through two views.