Amazon Quick: new agent job? Often it's a trigger, not a new agent

Three Amazon Quick engineers published a post on the AWS Machine Learning Blog on 6 November 2026 setting out the vocabulary their team settled on for agent work. Their claim: any agentic task can be placed with four plain questions, When, What, Who and How. The post is vendor-published and describes Amazon's own product design; it reports no measurements.

The four parts

A trigger starts the work and carries the task, so it answers both When and What. The post names four kinds: a person's message, one agent delegating to another, a schedule such as 'run financial research every Monday at 9am', and an event such as a new email, a Slack mention, a file change or a webhook.

An agent is the Who. The authors define it as a persistent actor made of three things: instructions, capabilities and learned context. The underlying large language model is described as a swappable parameter, and conversation history as per-session state.

Capabilities are the How, and they run from flexible to fixed. Tools are single operations with a typed schema. Skills are structured guidance with room for judgment. Workflows are declared steps with branching and error handling, with decision logic in code. The post's example of a workflow is a purchase-approval route: budget check, manager sign-off, then finance, each step recorded.

What changes in a rollout

The authors say a request like 'monitor my Slack channel and send me a daily digest' is usually a trigger added to an existing agent. They list three valid builds: a scheduled one-line instruction, the same task backed by a skill with explicit formatting rules and channel priorities, or the task invoking a workflow. Their framing is that the choice is a 'confidence contract': a team that trusts the agent takes the first, one burned by inconsistency the second, and a compliance-sensitive process the third.

They also argue skills and workflows will outlast model improvements. Their reasons are organisational: skills capture and share institutional expertise, and workflows give auditability and step-by-step visibility. A skill can tighten into a workflow, or a workflow loosen into a skill, which they say Quick handles through a continuous learning framework to be described in a later post.

What the post leaves out

There is no data on cost, error rates or how often each build fails, and no customer example. The learning claim, that 'the tenth run is markedly better', is not backed by a number. The post is the first in a series; later posts will cover Quick's agentic operating system, background learning and proactive feeds.

For a team with a backlog of 'can the agent do X?' requests, the practical use is triage: ask what starts the work, which agent owns it, and how much of the path you want fixed in advance. Whether Amazon's boundaries match your tools is something only a pilot will show.