The capability equation

Model + Tools + Skills + Memory = Agent.

What turns a chatbot into an agent

A chatbot has a model. An agent has four things.

• Model — the reasoning engine. For the Anthropic stack you might use a large model like Opus for nuanced judgement, medium-sized Sonnet for the daily workload, and smaller Haiku for speed and lower cost.
• Tools — what the agent can DO. Query a database, call an API, search the web, send an email. Executed outside the model; deterministic and auditable.
• Skills — what the agent KNOWS that is specific to your domain. Structured expert knowledge, injected into context, that shapes how it reasons.
• Memory — what the agent REMEMBERS across turns, tasks, and sessions. Short-term, working, and long-term.

An agent's capability is the product of all four layers — not any one alone.

The thought experiments

• Tools without skills? An agent that can act, but doesn't know your rules. A new hire with system access but no training.
• Skills without memory? An agent that knows the playbook but forgets every conversation. Groundhog Day.
• Memory without governance? An agent that learns your patterns — including your shadow workarounds. Water finds a way.

Each of these failure modes is real, common, and avoidable. They are avoidable only if you build all four layers deliberately, in order, with governance layered on top.

How the rest of the Build module is organised

• Model selection — which tier for which task.
• Context windows — the shared budget every agent lives in.
• Tools and MCP — what the agent can DO, and how it connects to other datasources.
• Skills — what the agent KNOWS.
• Memory — what the agent REMEMBERS.
• System prompts — the persistent instruction that glues it all together.
• RAG and context — retrieval vs fine-tuning vs long context.
• Orchestration — single agents, sub-agents, multi-agent flows.

Evaluation and testing are the next pillar on their own — see /courses/agents/learn/evaluate.