Why build another harness?

Why build another harness when there are hundreds of them and a few new ones shipping every single day?

Everyone building a harness right now is chasing autonomy and we're all looking in the same forest but everyone is barking up a different tree. Cyboflow makes two core bets about which tree is the right one:

1. Jagged intelligence is here to stay: as long as AI means LLMs, we're going to be dealing with agents that are superhuman in some dimensions and worse than your average middle schooler in others
2. Workflows are key to solving the memento problem: the only way to work with a model that can have a PhD level intelligence yet forget what you told it 15 minutes ago is to break down tasks into steps that can fit within a context window

The whole product is built around those two bets.

Why do I think Jagged Intelligence is here to stay? If you've been following me, you know the answer is RLVR.

RLVR explains the seeming contradiction that you can have a model capable of cracking math problems that stood unsolved for decades and yet still can't string together two paragraphs that don't sound robotic.

RLVR doesn't just explain the jaggedness, it gives shape to it. The jaggedness isn't random, it maps to verifiability

If jagged intelligence is here to stay, the answer isn't to remove humans from the loop but instead to focus human attention on the areas the models are incapable of doing well. Because of RLVR, that's generally the places where there's no right answer.

But human attention is expensive, so you have to treat it like your scarcest resource. If a task is verifiable, hand it to the agents. If it isn't, that's exactly where your judgement is worth something. Don't spend it anywhere else.

The second core bet is on workflows and it stems from a simple fact. When you've got a context window limitation, the only viable solution is to break down a task into context window sized chunks.

Workflows also solve a second problem. How do you meld humans and agents into a single system? You designate certain steps as human ones.

But while "workflows" might serve as a broad brush answer, it belies the complexity underneath. The challenge with workflows is they're a fine balancing act.

Make the steps too small and you're repeating context. Too large and performance degrades. Pull humans in too frequently and you're wasting the scarce thing. Wait too long and the agents will have drifted way off course.

Both these bets give Cyboflow its name. Cybo from cyborg, because the goal is melding humans and machines instead of separating them. Flow from workflow, because that's how you actually do the melding.

They also drive Cyboflow's core product principle: maximize the value of human attention. Each incremental feature earns its place by making the most of eyeballs looking at it.

While everyone is chasing autonomy by seeing where they can pull humans out of the loop, Cyboflow inverts that and assumes they need to be there, not just as minders but as a fundamental part of how value gets created. And if humans are going to be part of the loop, then we need to focus our product attention on how to use that input as effectively as possible.

How you actually do that is where all the fun is.

Adapted from The Cyboflow thesis.