This is a knowledge site about the software around a language model: the harness. It starts from one HTTP call, adds a loop, names the eight jobs that the loop needs done, and then reads SDKs, coding harnesses, frameworks and managed platforms as different owners of those same jobs. It is written for engineers and architects who build agent systems or choose between products that claim to build them.
What it is and is not#
| It is | It is not |
|---|---|
| A map: one way to read any agent product, as eight jobs with an owner each | A standard, or the only way to divide the work |
| First principles, then abstractions: each problem in plain code before the product that hides it | A product catalogue or a buying guide |
| One running example, a property-insurance quote, followed through every chapter | A record of any real insurer; the company, the data and the vendors are invented or simulated |
| Product behavior read on the vendor's own documentation on 16 November 2026 | A promise that a product still behaves that way; check before you build |
| Descriptions by job: what a product runs for you and what stays with you | A ranking. No product is called best |
What is original here and what is not#
| Idea | Status |
|---|---|
| A model call, tool use, the agent loop | Documented behavior of model APIs |
| The eight jobs of a harness | A checklist proposed on this site. Other authors divide the same work differently |
| The six-rung ladder | A way of ordering existing kinds of product, proposed on this site |
| The words BUILT IN, INTEGRATES, YOU BUILD, MANAGED, NOT ITS JOB | This site's reading of each vendor's documentation, not a vendor's own label |
| Checkpoints, retries, idempotent side effects, outbox, traces, golden sets | Standard engineering practice, applied to agents |
| The miniature harness | Written for this site, to be read in one sitting. It leaves things out on purpose, and each page says what |
How the facts were checked#
| Kind of claim | How it was checked |
|---|---|
| Product behavior | Read on the vendor's own documentation on the day of writing and logged with the page address and the date. More than 500 claims are logged. Where a page did not say, the site says "the docs do not say" |
| The product by job map | Built from those readings. A cell is a reading of documentation, not a test of the product |
| The miniature | Run. Every command and output shown from it is real, against a scripted stand-in for the model. It never called a live model for this site |
| Code in Python, TypeScript and Java | Each block says how far it was checked: compiled against the real package, constructed without calling a model, run against a local stand-in, or documented shape only |
| The full build | Described from the author's own implementation. Its vendors are simulated and its history is seeded. Its numbers are quoted, not re-measured |
| Numbers | Illustrative unless a source is linked. Token counts from the miniature are the scripted model's estimates |
Documentation changes. Several facts on this site differ from what the same pages said two days earlier, and the pages say so where it matters. Treat every product claim here as true on the day it was read.
What the miniature leaves out#
The miniature is about 450 lines so that it can be read whole. A production harness needs more. The gaps are stated on the pages where they matter, and collected on Build your own. Among them: it does not validate tool arguments against the schema, it screens untrusted text with four simple patterns, its trace does not record every token type, and what it calls an outbox is a key store without a worker.
Who wrote it#
Sam Muthu, an engineer and architect who builds agent systems and holds Anthropic's Claude Certified Architect credentials (Foundations and Professional). More of his work is at sammuthu.com. A sister publication, Diff to Deploy, follows one code change through a delivery pipeline that uses coding agents.
The site is a working version. If a claim is wrong or out of date, the fix is to read the vendor's page again and correct the sentence.