Amazon described a taxonomy of primitives for building AI agents in the Amazon Quick product
Amazon published on the AWS blog a taxonomy of four primitives (trigger, agent, capabilities, memory) for building AI agents in the Amazon Quick product, based on the questions When, What, Who, and How.
Amazon described on the AWS Machine Learning blog a taxonomy of primitives it used in developing the Amazon Quick product. This is the first part of a planned series of posts. According to the company, when building agents, a problem kept recurring where the team had no shared definition of the basic building blocks of agentic work, and the same concepts were renamed under different names. The resulting framework breaks down agentic work into four questions: When, What, Who, and How.
The trigger represents the activation mechanism and at the same time carries the instructions for the task — it is therefore the answer to both When and What at once. According to the company, the type of trigger affects the agent's behavior: a scheduled background task behaves more conservatively, while a task assigned by a human responds more flexibly. The agent is defined as a persistent, learning actor answering the question Who — it is formed by instructions, capabilities, and history, not by a specific large language model, which the company describes as an interchangeable parameter. Capabilities answer the question How and include tools (atomic operations with a typed interface), skills (structured guides with room for judgment, loaded progressively), and workflows (deterministic steps with a fixed order and logic in code).
According to the company, this structure makes it possible to merge chat-based, scheduled, and "autonomous" agents into a single actor that can be conversed with, that can be scheduled, and that can coordinate and even examine its own past work. The company also describes a mechanism whereby a well-established skill can be refined into a workflow ("crystallization"), and conversely a workflow that runs into edge cases can be released back into a skill ("dissolution") — according to the company, this mechanism is to be described in a future post about continual learning.
The source article is incomplete; the remaining part of the text was not available. See the source article for details.
Why it matters
The framework addresses a specific design problem — how to distinguish whether a user request is a new agent, a scheduled instruction to an existing agent, a skill, or a workflow. For teams developing their own agent systems, this is a possible pattern for structuring agent architecture without having to reinvent the same concepts under different names.
Two audiences, two different impacts
What this means
For individuals
Developers and architects designing AI agents get a concrete conceptual framework for distinguishing whether a user request is a new agent, merely a new trigger, a skill, or a workflow.
For a business
Companies building their own agent platforms gain a reference vocabulary (trigger, agent, capabilities, memory) that can reduce duplicate reinvention of the same concepts under different names when designing a product.
Development More business impacts →Check the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.