Amazon Quick publishes a taxonomy of agentic primitives

The team behind Amazon Quick says its agents moved from answering questions to taking actions such as scheduling meetings and generating reports, and that the same concepts kept being rebuilt under different names. The post splits agentic work into four questions, with triggers covering When and What, agents covering Who, and capabilities covering How. Triggers can be human-initiated, agent-delegated, scheduled or event-driven.

An agent is defined as a persistent, learning actor made of instructions, capabilities and learned context, while the underlying LLM is treated as a swappable model parameter. Capabilities include tools, which are atomic operations with typed schemas, skills, which give the agent guidance it loads progressively, and workflows, which run declared steps in a fixed order. According to the post, a skill can be tightened into a workflow, and a workflow that hits many edge cases can be loosened back into a skill.

The post says a request to monitor a Slack channel and send a daily digest is usually a scheduled trigger added to an existing agent rather than a new agent. The same request can be handled with a one-line instruction, a skill or a workflow, and the right choice depends on how much latitude the user wants the agent to keep. This is the first post in a series that will cover the agentic operating system, background learning and proactive feeds.

Builders can treat a recurring request as a trigger on an existing agent, reserving skills for reusable expertise and workflows for steps that need fixed ordering and audit trails. The post argues these choices hold their value as models improve, since skills capture institutional knowledge and workflows provide auditability. Because agents accumulate learned context across runs, developers should plan for that state when they design agents.