The Most Useful Sentence Is the One He Buried at the End

Tony Kipkemboi's thread on agent frameworks vs. agent harnesses is worth reading if you've been nodding along to these terms without quite knowing where one ends and the other begins. He's got the right map. He draws the wrong conclusion from it.

The piece earns its keep with one specific move: the LangChain breakdown. Framework, runtime, harness — three distinct layers, one company, named concretely. That's not vocabulary for its own sake. It shows you how the spectrum actually plays out in a real product stack, not just in theory.

Everything else builds toward a single line: "Frameworks are for builders. Harnesses are for users."

That's where it falls apart. The line is clean enough to screenshot and vague enough to mean almost nothing. Most of the people choosing between these tools aren't choosing between building and using — they're builders who need to ship, and they need to know which decisions the harness has already made on their behalf. Accepting a harness isn't just trading control for speed. It's accepting someone else's assumptions about your context, your failure modes, your data environment. The thread doesn't make you ask: whose opinions, exactly? It should.

There's also the small matter of what happens when the harness breaks. A framework forces you to understand the pieces. A harness insulates you from them — right up until something goes wrong and you have no conceptual scaffolding to debug it.

The paragraph worth the most attention comes last, after the piece has already summarised itself:

"Sometimes — really most of the time — you need to bypass the agent frameworks entirely and build a simple ReAct agent using the model endpoints directly."

That's the practitioner reality check. Everything before it is taxonomy. That sentence is what you'd say to someone actually shipping something.

He built a neat spectrum. Then admitted the spectrum might not be the point.