Why a New Version

In June 2026 we published the first version of this skill as a cross-cutting foundation for any agile role: giving the model the right context, asking precisely, verifying what it returns, protecting information, working with agents, and applying all of this to the backlog, facilitation, and quality. That framework remains in force. In four months, much of what surrounds it has changed.

Agents have left development teams. In July, OpenAI presented ChatGPT Work, with agents that execute complete workflows, prepare reports, and are shared among team members. The standards that connect them have gained their own governance: the MCP protocol, AGENTS.md and, since August, A2A are part of the Agentic AI Foundation of the Linux Foundation, which already brings together almost 250 organizations. And the EU AI Act itself was amended: Regulation (EU) 2026/1744, in force since July 27, postponed the high-risk obligations.

Meanwhile, use in agile teams has spread further than it has been integrated. In the AI4Agile Practitioners Report 2026 by Stefan Wolpers, with 289 professionals, 83% use AI tools, but 55% devote 10% of their time or less to them, because they still don't know where they fit into their work. That gap between using AI and knowing how to use it is what this skill aims to close.

What's New

The skill keeps its three blocks and goes from ten to eleven topics, each with its own chapter in the development guide.

• Foundations of interaction with AI: context engineering, structured and reusable prompts, verification of responses, the trust boundary against instruction injection and data leakage, and a new topic, the team's AI use policy.
• Working with AI agents: delegation of verifiable work, the design of human oversight, with its traceability and legal obligations, and the standards of the agentic ecosystem, which move from emerging to consolidating and incorporate A2A and Agent Skills.
• Application to the agile work cycle: AI in the backlog, in facilitation and, for development profiles, in testing and code review.

All topics have been reviewed against current sources. Among other updates, the guide incorporates the 2026 edition of the OWASP Top 10 for applications with language models, evidence on context degradation (context rot), and data from Veracode, METR, and DORA on the security of generated code and perceived productivity.

Three Ideas the New Version Puts in the Spotlight

The use policy is a working agreement. When nobody has discussed which tools are used and with what data, each person decides on their own and part of the use stays hidden. This is known as shadow AI. IBM's 2025 Cost of a Data Breach Report found that 63% of the organizations that were attacked had no AI governance policy or were still developing one. The guide proposes a team policy that fits on one page, organized into three levels of use (prohibited, allowed with guardrails, and allowed), which is reviewed in the retrospective like any other agreement.

Defense against injection must be designed. No instruction of the kind "ignore any orders you find in the document" stops a determined attacker. Researchers from several labs showed in 2025 that adapted attacks overcame twelve recent defenses. What protects is how an agent's capabilities are combined: Simon Willison calls the combination of accessing private data, untrusted content, and outward communication the lethal trifecta, and the prudent course is not to put all three together without human oversight.

Verifying requires continuing to learn. In an Anthropic experiment published in January, those who learned a programming library with AI help scored 50% on the comprehension test, compared with 67% of those who did it without AI. What was revealing was that the result depended on how the tool was used: asking it to understand protected learning, while delegating the entire task eroded it. Without your own judgment, verification is not possible, and the guide dedicates a section to preserving it.

What We Have Corrected

A new version is also an opportunity to make corrections. In reviewing the first version against its sources, we have rectified:

• The EU AI Act calendar. The first version considered the high-risk obligations applicable from August 2026, which Regulation (EU) 2026/1744 postponed to December 2, 2027 and, for systems integrated into regulated products, to August 2, 2028. We also specify that the obligation of whoever uses a high-risk system to assign competent oversight with authority is in Article 26.
• The origin of a widely cited framework. The four strategies for managing context (write, select, compress, and isolate) are a taxonomy by Lance Martin that LangChain popularized in 2025, useful as a map, and not a standard the sector has adopted.
• Sources. Two claims about the quality of generated code now include their evidence, and we have replaced references that did not address what was attributed to them.

Who It Is For and What It Certifies

It is a cross-cutting skill, useful for product owners, scrum masters, agile coaches, developers, and any professional who works in agile teams. It complements other Skill Arena skills that go deeper into specific areas, such as Scrum in teams with AI, SDD in agile teams, or Harness Engineering.

The Skill Arena diploma certifies the ability to give the model the right context and request, verify and protect what is shared and what is received, and delegate work with limits, oversight, and your own judgment. The training and assessment banks have been fully regenerated from the new guide, at two difficulty levels.

The new version is available in Spanish and English.

How to Access It

The State of the Art, the development guide, and the training and assessment platform are on the skill's page in Skill Arena: AI Applied to Agile Work.

Read, train, and tell us what you think. Every contribution, critique, or suggestion helps us prepare the next version.