What they are, how they work, and why they matter now
Most software waits for you. You open it, click, it responds. AI agents are different. An agent is software you can give a job to — not just a question — and it goes away and does it. It reads, decides, acts, and reports back. Less like a calculator, more like a capable colleague who works at machine speed.
That shift — from tool you operate to colleague you delegate to — is what makes this moment worth paying attention to.
Every agent, simple or complex, is built from the same six layers stacked on top of each other. The diagram below shows all six — and what lives inside each one.
AI Agent · Functional Architecture · Six Layers
At the bottom is the reasoning engine — the AI model itself. This is the brain. It reads its instructions, looks at the task in front of it, and decides what to do next. The model alone is just a thinker. It needs the layers above it to give it a job, hand it tools, and keep it working over time.
The next layer up is the harness — the platform that keeps the agent running. This turns a single AI response into an ongoing workflow. It schedules tasks, handles errors, and can split a large job across multiple specialist agents working in parallel. Claude CoWork sits at this layer.
Above that are skills and memory. A skill is a written set of instructions that tells the agent exactly how to handle a specific type of task — your process, your rules, your context. Memory lets the agent carry knowledge between sessions so it does not start from scratch every time. Skills are what make an agent feel genuinely expert. Anyone can have a capable agent; skills are what make it your agent.
Then come connectors and MCP — the wires that plug the agent into the tools and systems you already use. Email, calendars, CRMs, finance systems, document libraries. MCP, a universal standard now adopted across the industry, has made connecting to almost any system fast and straightforward. Skills tell the agent what to do. Connectors give it somewhere to do it.
The action layer defines what the agent can physically do. The obvious ones are sending messages and updating records. The more powerful capabilities are browser use — the agent navigating any website and interacting with it just as a person would — and computer use, where the agent operates a desktop application directly, even one with no technical integration available. These mean that almost no system is unreachable, including legacy ones everyone assumed would never be automated.
At the top sits governance — the controls that keep humans appropriately in the loop. Rules about when to stop and ask rather than proceed. Logs of every decision made. Checks that flag when the agent drifts from how it was designed. This layer is usually built last and wished for first.
These three building blocks are what take an agent from interesting to genuinely useful — and together they give you the reach to handle almost any real-world workflow.
Write a skill, add a connector, spawn a sub-agent. Between those three moves, there are very few business workflows that cannot be meaningfully automated.
Skills encode your specific way of working — your rules, your judgement calls, your context. A well-designed skill is the difference between an agent that does the job and one that does your job.
SKILL.mdYour processYour rulesConnectors reach into every system you already use. MCP has made this a configuration task rather than an engineering project. For anything without an API, browser and computer use bridge the gap.
MCPAny systemBrowser useComplex jobs get broken down and handed to specialist agents coordinated by an orchestrator. A team of focused agents can cover an astonishing amount of ground — more than any single agent working alone.
Multi-agentOrchestrationScaleThe practical implication is significant. There are very few business workflows that cannot be meaningfully covered using some combination of these three things. The question is no longer whether it is technically possible. It is whether you have designed the agent well enough to trust it with the work.
You do not need to be a developer. The reasoning engine is already built. The infrastructure is already managed. What you are designing is the judgement — what the agent should know, what it is allowed to touch, and when it should stop and ask a human. That is a thinking problem, not a coding problem. And it is one that professionals who understand their own processes are uniquely well placed to solve.
The organisations building that fluency now will set the pace. The tools are ready.
A 3.5-hour hands-on session where you explore the magic, possibilities and limitations of Claude Agents.
Programme
What is an AI agent? Core concepts — Harness, Skills, Plugins, MCP — in plain English. A live demo shows Claude CoWork autonomously navigating a computer and solving a real problem without human input.
Set up and launch your first agent using Claude CoWork. Give it a real-world task — read a folder of messy documents, organise the information, and write a clean summary, all on its own.
Give your agent a precise, repeatable capability. Upload a document and build instructions that force the agent to extract exactly the information you need, every time.
Give your agent real tools to interact with outside services via MCP — query a live database, read documents from cloud storage, or search the web on your behalf.
Discover Claude Code — an AI that reads, writes, and organises files on your computer using plain English. See how it fits alongside CoWork in your day-to-day workflow.
Step away from your screen, grab a drink, and connect with fellow attendees.
Put it all together and build an agent you'll actually use after today. Choose your own use case — Email Summariser, Meeting Notes Assistant, Document Analyst. Instructors circulate to help.
Share your screen and show how your agent planned, acted, and delivered results. Get feedback, ask anything, and leave with a Prompt Cheat Sheet to keep building this week.
What you need on the day