Concepts

Tool-independent explanations of the ideas every agent system is built on. Read these before picking a framework.

How the pages fit together

An agent is a model plus a harness. The harness is the software around the model: prompt, tools, state, limits, observability. At the core of every harness is a loop: call the model, run the tool it asked for, feed the result back, repeat. When one loop is not enough, the loop becomes a graph: several steps with different prompts and tools, connected by edges that decide what runs next.

Pages

• Agent — What an agent is, what the model does alone, and what the harness adds. Why the same model becomes a different agent in a different harness.
• Harness — The parts of a harness, what each one is for, and how to match a harness to a task.
• Loop — The model-call-tool-repeat cycle, its stop conditions, and the ways it fails.
• Graph — Multi-step control flow: nodes, edges, shared state, and when a plain loop is no longer enough.
• Tools — Designing tool interfaces the model can call reliably: names, schemas, results, granularity, side effects.
• Context and Memory — What goes into the prompt at each step, budgeting and ordering, managing history, retrieval, and memory.
• Stacking Loops — Four loops around an agent: agent, verification, event, improvement. What each adds and where humans belong.

Coming soon

• Middleware and hooks — Running code before and after each model step
• Evals — Measuring whether a change to the harness made the agent better

Related pages

• AI Agents and Memory covers memory in depth.
• The RPI Loop is a workflow for humans using coding agents, not the agent's own loop.
• LLM Rubik's Cube Race is used as a running example: four models, two harnesses, one task.