Skip to content
CRAFTDEXA

Theme guide: AI

Applied AI, agents and evaluation.

Most AI projects stall between a convincing demo and a system someone is willing to rely on. What closes that gap is rarely the model: it is a narrow task, the tools an agent is allowed to use, a way of measuring results, and a plan for when it is wrong.

What is moving now

Dec 2025

MCP moves under the Linux Foundation

Anthropic donated the Model Context Protocol to the Agentic AI Foundation, a directed fund of the Linux Foundation that it co-founded with Block and OpenAI. For teams building agents, a common way to connect models to tools and data is now governed in the open rather than by one vendor.

SourceAnthropic announcement (opens in a new tab)

Dec 2025

OWASP publishes a Top 10 for agentic applications

The OWASP GenAI Security Project released its first Top 10 for Agentic Applications (2026 edition) on 9 December 2025: ten risks, ASI01 to ASI10, each with mitigations. It is a practical review list for any agent that can call tools or act for a user.

SourceOWASP GenAI Security Project (opens in a new tab)

May 2026

EU AI Act: high-risk deadlines pushed back (provisional)

Under the Digital Omnibus, negotiators reached a provisional agreement on 7 May 2026 to move stand-alone high-risk systems (Annex III) from 2 August 2026 to 2 December 2027, and AI built into regulated products (Annex I) to 2 August 2028. Confirm the final published text on EUR-Lex before planning around the dates.

SourceGibson Dunn briefing (opens in a new tab)

Summarised from public sources as of November 2026. Dates and details change, so follow the link to the source before relying on one.

A practical checklist

  1. 01

    Define the task before the model

    Write down the one job, who it serves and what a correct result looks like. If you cannot say it, you cannot test it.

  2. 02

    Build the evaluation set first

    Collect real examples, including the awkward ones, and run every change against them so progress is measured, not felt.

  3. 03

    Give an agent the fewest tools it needs

    Every tool is a permission. Scope credentials narrowly and make anything destructive wait for a person's confirmation.

  4. 04

    Treat all input as untrusted

    Documents, web pages and messages can carry instructions. Keep retrieved content apart from instructions, and check outputs before they act.

  5. 05

    Design the hand-off to a person

    Decide what the system does when it is unsure, and make that path fast. Escalation is a feature.

  6. 06

    Log enough to explain a decision

    Record inputs, tool calls and outputs, so a wrong answer can be traced, fixed and added to the evaluation set.

Where it shows up in our work

Related capabilitiesAI Integration ·Voice Agents ·Agentic Workflows ·Intelligent Automation ·LLM Applications ·Workflow Orchestration

Working on something in ai? Tell us what needs to exist.