How skills and agents work in AWS Quick: Repeatable tasks like status

Amazon Web Services expanded the automation capabilities of AWS Quick by detailing how autonomous AI agents, connected enterprise data sources, and modular skills integrate across the service. As detailed in AWS Machine Learning Blog's report, the platform allows teams to convert repetitive engineering and operations tasks like drafting status reports, compiling release notes, and triaging tickets into automated workflows. The cloud provider framed AWS Quick as an AI-driven environment for data analysis, application creation, task execution, and deep workplace research.

This architectural breakdown details how AWS Quick handles agent orchestration, data governance, ReAct loops, and the open Agent Skills specification. Technical leads, software engineers, and cloud architects can use this framework to design scoped custom agents and distribute reusable team skills across their AWS organizations.

01Default Quick agent and scoped custom assistants

The architecture places the agent at the center of execution to process incoming natural language requests against connected tools and data assets. AWS Quick supplies a default general-purpose agent that operates without a fixed scope, allowing any user to query accessible spaces, knowledge bases, dashboards, connectors, and skills. Because the built-in agent determines necessary tools on the fly, users can launch general inquiries without manual pre-configuration.

Organizations can also build custom agents configured for dedicated responsibilities, such as an employee onboarding assistant restricted to human resources spaces. Builders define distinct instructions for custom assistants while strictly bounding which spaces, connectors, and skills each agent can access. Every agent operates strictly within the security boundaries and resource permissions of the invoking user, preventing unapproved data retrieval across enterprise perimeters.

Once administrators test and publish a custom assistant, they can share the configuration with specific users or teams inside the AWS Quick account. These agents can run interactively within user chat sessions or execute independently in the background on scheduled routines.

02Curated spaces, knowledge bases, and connectors

Agents ground their responses using curated enterprise information gathered into collections known as spaces. Users and project groups upload files, folders, action connectors, datasets, and knowledge bases into a designated space to provide dedicated context for conversations. Attaching a space to a chat thread or agent profile restricts data ingestion to relevant business domains.

Quick syncs organized knowledge bases on automated schedules from external platforms including Google Drive, Microsoft OneDrive, Confluence, Microsoft SharePoint, Amazon S3, web crawlers, and Amazon Bedrock Managed Knowledge Bases. For structured metrics, AWS Quick Sight provisions datasets and dashboards directly through the AWS Management Console. These datasets establish semantic metadata layers over relational database tables, enabling natural language queries over complex analytics.

Connectors provide the execution bridge between AWS Quick agents and native enterprise applications. Cloud connectors maintain continuous authentication across Microsoft Outlook, Slack, Jira, Asana, Salesforce, and remote Model Context Protocol servers on any user device. Local connectors run on a user workstation to access locally hosted Model Context Protocol servers whenever the local machine remains powered and connected.

03ReAct loop execution and bundled system skills

Every skill represents a reusable instruction package teaching an agent how to execute a specialized workflow using available tool interfaces. Agents in AWS Quick execute through a managed agent harness structured around the reasoning and acting ReAct loop pattern. This model-driven design mirrors open orchestration architectures like the Strands Agents SDK to iteratively reason over prompts, invoke connectors, evaluate intermediate results, and generate responses.

The platform includes immutable system skills that run foundational actions without requiring third-party plugins. Built-in system skills can generate Word documents, PowerPoint presentations, Excel spreadsheets, and PDF deliverables using bundled execution scripts. Other native skills conduct web searches, execute arbitrary code, transcribe audio and video, synthesize images, manage persistent knowledge graphs, and dispatch programming tasks to the Kiro coding agent.

System skills also control internal administrative functions like scheduling routine runs, configuring AWS Quick Flows, and logging personal tasks. Workspace users can toggle individual system skills on or off within their environment, but the system prevents editing, duplicating, exporting, publishing, sharing, or deleting these default skills.

04Agent Skills specification and folder structure

AWS Quick adopts the open Agent Skills specification originally released by Anthropic to maintain standardized skill packaging across industry agents. Following the desktop preview launched in April 2026, AWS Quick on desktop reached general availability in September 2026 using this specification. Each skill lives inside a dedicated directory containing a mandatory root SKILL.md file alongside optional reference documents and executable scripts.

| Specification Parameter | Requirement or Milestone | | Skill name length cap | 64 characters | | Skill description length cap | 1,024 characters | | Skill instruction length guideline | Under 5,000 tokens | | Desktop client preview | April 2026 | | Desktop client general availability | September 2026 |

The frontmatter located between triple-dash delimiters defines the skill name and description, which govern discovery during runtime. The runtime relies on progressive disclosure, loading only the metadata into the initial context window until user intent triggers the skill instructions. The standard caps name length at 64 characters, limits descriptions to 1,024 characters, and recommends keeping instruction bodies under 5,000 tokens.

05Chat creation prompt and team skill sharing

Builders can generate new skills directly through conversational chat, upload prebuilt Markdown files, or import packaged zip archives containing a root SKILL.md file. When crafting a skill through chat, builders submit structured prompts detailing execution schedules, query sources, data transformations, and target output formats.

Every <cadence>, I have to grab data from the following sources:
- <List sources and how to query them>
And do <describe data transformations and exact sequence of steps>.
Once finished, create a report <describe format or provide template>.

The underlying system compiles this natural language description into a structured skill directory, pairing the SKILL.md instructions with executable Python calculation scripts or JavaScript rendering routines. Once published, custom skills become discoverable in the organizational catalog, allowing teams to automate Jira sprint summaries, Slack announcements, and cross-application handoffs.