Taxonomy of agentic primitives in Amazon Quick
· Engineer's Notes · Cem Koyluoglu
A concise overview of Amazon Quick's four-question taxonomy—When, What, Who, How—and its impact on production reliability and design.
What happened
The Amazon Quick team identified a recurring lack of shared terminology for building agentic workflows. To address this, they defined a taxonomy based on four plain questions: When?, What?, Who?, and How?. The taxonomy separates work into three primitive categories:
- Triggers – the activation mechanism that also carries the task description. Four trigger types are described: human‑initiated (interactive, synchronous), agent‑delegated (programmatic, potentially recursive), scheduled (time‑based, deterministic), and event‑driven (external signals such as emails or Slack mentions). The trigger type influences execution behavior, for example a scheduled task may stay conservative while a human‑initiated one remains responsive.
- Agents – persistent, learning actors that answer the Who? question. An agent comprises three components: Instructions (persona, role, operating principles), Capabilities (tools, skills, workflows that can be added or removed dynamically), and Learned context (memory of past interactions that improves performance over successive runs). The underlying large language model, conversation history, and triggers are considered extrinsic to the agent.
- Capabilities – the How? of execution, spanning a confidence contract from fully agentic to fully deterministic. Three capability forms are defined: Tools (atomic operations with typed schemas), Skills (structured guidance with room for judgment, loaded on demand), and Workflows (deterministic coordination with fixed steps and code‑driven decision logic). Skills and workflows can evolve into each other based on usage patterns.
The post illustrates these primitives with the example “monitor my Slack channel and send me a daily digest,” showing that the same user request can be satisfied by a scheduled trigger with a simple instruction, a skill with explicit formatting rules, or a workflow with deterministic error handling. Learning across executions refines the agent’s performance without manual re‑coding.
Why it matters in production
The taxonomy provides a clear abstraction layer that can improve system reliability. By classifying triggers, teams can enforce appropriate safety policies—e.g., scheduled triggers may default to conservative actions when confidence is low, reducing the risk of unintended side effects. Distinguishing capabilities into tools, skills, and workflows enables deterministic audit trails for compliance‑sensitive processes, which is essential for regulated environments.
Memory‑based learning reduces operational cost over time: as agents accumulate learned context, fewer explicit instructions are needed, lowering the volume of human‑in‑the‑loop interventions and associated latency. The ability to swap the underlying LLM without altering agent definitions supports model upgrades without breaking existing integrations, preserving stability while benefiting from performance gains.
Security implications are clarified through the explicit trigger model. Event‑driven triggers can be gated by authentication checks, while agent‑delegated triggers allow fine‑grained permission controls between agents, limiting privilege escalation pathways.
Engineering takeaways
- Adopt the four‑question taxonomy to standardize design discussions and reduce duplicated terminology across teams.
- Classify each entry point as a trigger type and enforce default execution policies (conservative vs. responsive) based on that classification.
- Encapsulate reusable logic as skills when judgment is required, and as workflows when deterministic coordination and auditability are needed.
- Leverage the agent’s learned context to iteratively improve performance, monitoring for diminishing returns to decide when to transition a skill into a workflow.
- Separate the LLM layer from agent definitions to enable seamless model upgrades without re‑engineering existing primitives.
Related work on this site
- Agentic & LLM Automation — Multi-step LLM pipelines that run unattended: model cascades with fallbacks, validation gates, scheduled automation and alerting when something breaks.
- RAG & Grounded LLM Systems — LLM features that answer from the right source instead of from memory: document-grounded assistants, transcript-grounded chat and quality gates around generated output.
- YouTube AI Summarizer — Open-source Chrome extension on the Chrome Web Store: AI summaries, key points, deep analysis, a two-host AI podcast (Gemini TTS) and transcript-grounded chat for any YouTube video, using Groq or Ollama Cloud with the user’s own key.
- Automated AI News Pipeline (this site) — The pipeline behind this site’s Tech News: scheduled GitHub Actions scrape sources, an LLM cascade on Groq rewrites and enhances articles, and layered quality gates (date integrity, language checks, instruction-leak detection, duplicate detection) decide what gets published, with Telegram alerts.
This note was drafted with AI assistance from the primary source credited on this page, and automatically checked against that source before publishing.
Source: https://aws.amazon.com/blogs/machine-learning/taxonomy-of-agentic-primitives-in-amazon-quick/