Agents & Harness
What an AI agent is, what a harness does, and how the pieces of an AI environment — context, tools, sandbox, memory — fit together.
An LLM is a brain in a jar. An agent is that brain given hands: tools it can use, a loop of acting and thinking, and guardrails around it all.
This section explains the machinery — the agent loop, the harness, the environment, and who builds what.
One step of an agent: think → act → observe — then loop until the job is done.
What is an AI agent?
A model plus a loop plus tools: the act–think–observe cycle that turns a chatbot into a worker.
What is a harness? (And why it matters)
The harness is everything around the model: prompts, tools, memory, safety. Same model + different harness = different product.
The AI environment: context, tools, sandbox, memory
The four rooms the agent lives in — and which of them you can trust.
System prompts, in plain English
The invisible instruction sheet at the top of every agent conversation — what it contains, why it's long, and how to read one.
Function calling / tool use
How the model actually 'uses' a tool: it doesn't — it proposes a JSON call, the harness executes it. Plus the three classic ways it goes wrong.
Multi-agent systems
When one agent is not enough: the three common patterns, the honest case for staying single, and the cost multiplier nobody mentions.
Agent security: prompt injection & guardrails
Why 'ignore your instructions' hidden in a PDF can work, and the five guardrails that actually hold — the model is the least secure layer in the stack.
Who builds what? (The layer map)
Model labs, API infra, harness vendors, and you — four layers, four different jobs, and why 'the model is bad' is usually the wrong diagnosis.