Amazon Web Services (AWS) has introduced a foundational taxonomy aimed at standardizing the building blocks of agentic artificial intelligence systems, as developed for its Amazon Quick platform. This framework, built around four core questions – When?, What?, Who?, and How? – seeks to provide a common vocabulary for developers and users engaged with AI agents capable of performing complex actions.
The Chenab Times has learned that the taxonomy emerged from the development of Amazon Quick, where the AWS team observed recurring concepts underpinning agentic tasks. These included the entity performing the work, the signal that initiates the task, standardized procedures, and mechanisms for retaining learned information. The lack of a unified definition led to confusion and redundant development, particularly in how user requests were interpreted and translated into agent actions.
Defining the Agentic Framework
The newly defined framework breaks down agentic work into four key primitives:
- Triggers (When and What): These initiate an agent’s work and carry the initial task description. Triggers can be human-initiated (conversational requests), agent-delegated (one agent invoking another), scheduled (time-based execution), or event-driven (signals from external systems like emails or file changes). The type of trigger influences the agent’s behavior, with human interactions demanding responsiveness and scheduled tasks allowing for more conservative execution.
- Agents (Who): Defined as persistent, learning actors, agents are the core entities that interpret instructions and execute tasks. An agent comprises its standing instructions (persona, role, context files), available capabilities (tools, skills, workflows), and accumulated learned context or memory. This definition emphasizes that the agent is the interpreter of instructions, not the instructions themselves, allowing for distinct behaviors even with identical inputs due to differing histories and learned contexts.
- Capabilities (How): These represent the methods an agent employs to perform actions, ranging from fully deterministic to requiring agent judgment. Capabilities are further categorized into:
- Tools: Atomic operations with defined schemas, such as querying a database or sending an email.
- Skills: Structured guidance with room for interpretation, bundling instructions with resources to describe an approach rather than a rigid sequence.
- Workflows: Deterministic coordination of steps with fixed execution order, control flow, and error handling, typically implemented in code.
The framework emphasizes that most of these primitives remain invisible to end-users, aiming to simplify the interaction with AI agents. However, a shared vocabulary is intended to make the underlying system more comprehensible for those building upon it.
Addressing User Ambiguity
AWS highlighted the ambiguity often present in user requests. For instance, the command “monitor my Slack channel and send me a daily digest” could be interpreted as a new agent, a scheduled instruction, a workflow, or a skill. The new taxonomy aims to resolve this by categorizing such requests more precisely, often defining them as a trigger added to an existing agent rather than a wholly new entity.
The value proposition of this approach lies in its ability to unify different types of agent interactions. A single agent can potentially handle conversational requests, scheduled tasks, and orchestrated operations, consolidating the concept of autonomous agents with those that are directly interacted with. This allows for a more coherent user experience where complex tasks can be initiated through simple commands.
The Role of Skills and Workflows
AWS posits that skills and workflows will retain their significance even as large language models improve. Skills serve as a mechanism for capturing, versioning, and sharing institutional expertise, while workflows provide auditability, step-wise visibility, and deterministic coordination across multiple services. The system also supports dynamic evolution, where skills can become more structured workflows as understanding deepens, or workflows can loosen into skills when encountering frequent edge cases.
This continuous learning framework, which will be detailed in future posts, allows tasks to evolve. A simple request can accumulate domain-specific learnings over time, improving subsequent executions without manual intervention. The system is designed to offer the simplest option that meets a user’s confidence requirements, with the flexibility to graduate to more autonomous operations powered by learning.
Practical Implications for Developers
The framework enables multiple valid approaches to fulfilling a single user request, differing primarily in the confidence contract established. A user seeking simplicity might opt for a scheduled task with minimal instructions, relying on the agent’s learned capabilities. Conversely, a user requiring strict control or auditability for compliance-sensitive processes might prefer a workflow with deterministic steps and error handling. This flexibility allows developers to tailor agent behavior to specific user needs and confidence levels.
The taxonomy, consisting of four fundamental questions—When, What, Who, and How—is designed to encompass any form of agentic work. By providing a universal vocabulary, AWS aims to streamline the development of agentic AI systems and abstract away underlying complexity for users, allowing them to focus on their desired outcomes rather than the intricate mechanics of AI execution.
Upcoming posts in the series are expected to delve into the agentic operating system that underpins Amazon Quick and its capabilities.
❤️ Support Independent Journalism
Your contribution keeps our reporting free, fearless, and accessible to everyone.
Or make a one-time donation
Secure via Razorpay • 12 monthly payments • Cancel anytime before next cycle
(We don't allow anyone to copy content. For Copyright or Use of Content related questions, visit here.)
The Chenab Times News Desk
