Taxonomy of agentic primitives in Amazon Quick

Judgment

Amazon Quick's platform and product teams now have to decide whether the four-question taxonomy becomes a binding design contract for new agent features or just documentation; the tradeoff is fewer rebuilt primitives against slower, more constrained feature design.

Basis

Thomas Delteil, Adi Kalyanpur, and Ron Litman write that as Quick's agents moved from answering to acting—scheduling meetings, running research, generating reports—the same primitives kept being rebuilt under different names: an actor, a start signal, a packaged procedure, and a memory. They converged on a taxonomy that decomposes agentic work into four questions—When?, What?, Who?, and How?—and use the request “monitor my Slack channel and send me a daily digest” to show how, without shared vocabulary, a user's single request can look like a new agent, a scheduled instruction, a workflow, or a skill. The mechanism is reuse through shared definitions: most of the framework stays invisible to end users, but builders beneath it get one vocabulary for reasoning about triggers, agents, capabilities, and learned context.