AI Agents

An agent's brain is an LLM — you must master calling it and shaping its output before building any loop on top.

An agent's whole behaviour is set by its system prompt and reasoning style — a bad prompt means a lost agent.

Without tools an agent only talks — tools are what let it search, compute, and change the world outside text.

Frameworks give you the agent loop, tool management, and memory ready-made so you focus on your task logic, not plumbing.

An agent that forgets every step is useless — memory lets it resume where it stopped and learn from context.

For an agent to act on your knowledge rather than its guesses, it must retrieve the right information when needed.

Big tasks fail if the agent jumps straight in — planning breaks them into executable, reviewable steps.

A single agent gets lost on complex work — splitting it across specialised agents under a coordinator raises quality.

The MCP protocol standardises how agents connect to external tools and data, so you reuse instead of rewrite.

Agents are non-deterministic and fail in strange ways — without tracing and evaluation you can't trust them.

An agent with power to act can do real harm — guardrails and permissions are your last line of defence.

An agent in a notebook is one thing; an agent serving users reliably at sane cost is another entirely.