Advanced Agent Building: Skills, Tools and MCP

Level

Advanced

Format

Live and instructor-led, remote or on-site

Length

Four live half-day sessions (3.5 hours each)

Audience

Solution builders, business technologists and agent owners who build agents for other teams

Cost

Free for customers and partners; quoted for other teams

Certificate

VDF AI Certified Agent Builder

Reply to applications

Within 2 business days

Labs

One after every module

6 modules and a capstone

The modules run across the four sessions. Each one ends with a lab in VDF AI.

1. Authoring and validating Skills

   How a Skill is structured, how it enters an agent’s context and how to get a bundle through validation.

   • The Skill directory: SKILL.md, references, assets, scripts and evals
   • Metadata that drives discovery: name, description, owner, version and allowed-tools
   • Progressive disclosure: short listings first, full instructions and bundled files on demand
   • Writing one narrow procedure with its judgement calls, stop conditions and output shape
   • Resolving validation problems in one pass and keeping secrets out of a bundle

   Lab

   Write a private Skill in the editor for a checklist your team already follows, move the background material into references and validate the bundle until no problems remain.

   Outcome

   A validated private Skill with a precise description and a narrow procedure.

2. Binding, invoking and testing Skills

   Attaching Skills to agents and network nodes, then proving they activate when they should and act only through assigned tools.

   • The Tools & Skills builder step and its report of missing tools
   • Explicit invocation with a slash command versus model-initiated activation
   • Testing trigger precision, trigger recall, output quality and safety
   • Reading activation events: Skills available and activated, files read and context added
   • Skills on network nodes: inherited bindings, node overrides and the auto, listing and inline modes

   Lab

   Bind your Skill to an agent and run positive examples, near misses, an ambiguous request, a missing tool and a deliberately unsafe request in the Agent Playground, then classify each failure from the events.

   Outcome

   A test set that tells a triggering problem apart from a tool failure or a weak procedure.

3. Skill versions, import and export

   Preserving a known-good procedure and moving Skills between clients without losing control of what they contain.

   • Changes that call for a new version snapshot, and the Versions view
   • Exporting a Skill ZIP and the review to run before anyone else receives it
   • Importing a ZIP and clearing every problem validation reports
   • Agent Plugin packages: exporting Skills with optional MCP configuration, and what plugin import extracts
   • Portability: metadata support, tool-name mapping and re-running positive, negative and safety evals

   Lab

   Save a new version of your Skill, export it, check the archive for credentials and personal data, then import it through Add Skill in your lab environment and re-run your safety cases against the imported copy.

   Outcome

   A release routine for Skills in which each observed result maps to a saved version.

4. Tools, grants and connectors

   Deciding what an agent is allowed to do, and keeping that decision visible to the administrators who govern it.

   • The tool catalogue that Agents and Networks share, and its six tool families
   • Attached knowledge or a tool: choosing by how stable the source is and how often it is needed
   • Why a Skill never grants a tool and allowed-tools only documents the capability set
   • Workspace default tool sets, and tools an administrator locks behind an explicit grant
   • Registering a remote MCP server, testing the connection and reviewing the tools it exposes

   Lab

   Register an MCP server and review the tools it exposes, then audit an agent’s tools against its job, remove every tool the job does not need and note in the agent’s description why each remaining tool is attached.

   Outcome

   An agent with the smallest tool set that does its job and a written reason for each tool.

5. Visual agents and model choice

   Getting usable images, charts and mockups, and matching each agent to a model that suits its inputs, volume and constraints.

   • Image Generator prompts that name subject, composition, style and palette
   • Chart Generator prompts: clean data, the takeaway to show and an optional chart type
   • HTML Mockup Generator prompts: purpose, sections from top to bottom, style and audience
   • Letting VDF AI pick the model, and the four reasons to pin one instead
   • Image-capable models for agents that must read screenshots, charts or diagrams

   Lab

   Produce a hero image, a chart from a clean table and a clickable mockup for one internal proposal, improving each with single targeted refinements rather than fresh prompts.

   Outcome

   Visual outputs refined in place, and a stated reason behind each agent’s model setting.

6. Evaluating, sharing and publishing agents

   Proving an agent is dependable before widening its audience one step at a time.

   • Testing on the hardest real input and checking each citation against its source
   • Re-testing after a knowledge source is restructured or the model behind an agent changes
   • Direct sharing, workspace library publishing and embedding in another app or page
   • Narrowing knowledge and tools for embedded and customer-facing agents
   • The Runs view, naming each change to a shared agent and archiving its predecessor

   Lab

   Test an agent against its hardest recent input, share it with one teammate, then prepare its library entry with a clear name, description, tags, a pinned example run and a line on who it is for.

   Outcome

   An agent published with evidence that it works and a clear statement of whom it serves.