A language model keeps nothing between two calls. An agent is that call in a loop, and the harness is the software around the loop: it builds the context, runs the tools, keeps the state, enforces the rules and records what happened. This site starts from one HTTP call and climbs, one abstraction at a time, to SDKs, coding harnesses, frameworks and managed platforms, asking one question at each step: who runs this job?
Every agent product promises to “handle” context, tools, memory and safety. Underneath, each one runs the same loop around the same stateless call, and the work around that loop either happens in its code, in yours, or nowhere. Without a map of that work, a team cannot tell what a product actually does for it, or what it still has to build.
What a model is from the outside, the thirty lines that turn a call into an agent, and the six rungs of abstraction built over the same call.
Context, tools and MCP, retrieval, control flow and state, guardrails, reliability and cost, observability and evals, and many agents: one chapter per job.
Provider and agent SDKs, coding harnesses, frameworks and graph runtimes, managed platforms, each described by which jobs it runs and which stay with you.
A whole harness in one readable file, and the map of products against jobs to decide, job by job, what to build and what to adopt.