The Cyboflow thesis

Why launch 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 means autonomy can only come from focusing human attention, not replacing it. As long as we're working with LLM-based agents, jagged intelligence will only get more jagged. Agents will be increasingly superhuman along some dimensions and yet remain worse than your average middle schooler along others. The only way you increase autonomy when agents are jagged is by focusing human attention where agents are weakest.
2. Workflows are the natural solution to the Memento problem. The second core challenge of getting agents to perform effectively is the Memento problem. You can have a PhD-level intelligence capable of cracking a Fields Medal problem that can't remember what you told it fifteen minutes ago. Addressing that inherently requires workflows — breaking tasks down into a series of steps that can fit within a single context window. Workflows also provide the answer for how to meld humans with agents while still increasing autonomy.

These two bets form its foundation and also give Cyboflow its name. Cybo from Cyborg because the way to maximize agent capability is to meld them with humans not separate them and flow from workflow because that's the way to meld the two most effectively.

Jagged intelligence means autonomy can only come from focusing human attention, not replacing it

How is it possible that a model which can solve mathematical problems that have gone unsolved for decades still isn't capable of stringing together two paragraphs without sounding robotic?

It all comes back to RLVR. Some of the most dramatic gains in model performance over the last few years have come thanks to a specific training technique — reinforcement learning with verifiable rewards.

RLVR is a training technique that rewards models when they answer a question correctly. It's an incredibly powerful tool that's allowed us to turn a text-prediction model into one capable of performing complex tasks that require sophisticated reasoning.

But it comes with one key requirement: it needs definitive answers. It can only help with problems where an answer is either right or wrong. Math checks out or it doesn't. Code compiles and passes tests or it doesn't. When the reward isn't verifiable there's nothing clean to train against. You can't unit test whether a piece of writing is moving or whether an interface will feel obvious to someone seeing it for the first time.

Because RLVR has been behind some of the most dramatic gains, and because capability tracks verifiability, the jaggedness isn't random. It maps to what can be checked. But the flipside of that is progress has been much faster in verifiable fields while it's remained relatively stagnant in unverifiable ones. It explains the paradox of incredibly sophisticated mathematical reasoning alongside missing basic common sense: the first is verifiable, the second isn't.

The models will get better over time at the unverifiable fields through better data and training, but it will be slower and far less dramatic than the progress we've gotten used to in the verifiable ones.

Which leads to a natural answer. Instead of trying to take humans out of the loop, treat them as your scarcest resource. You need human insight to complement the robots. You shouldn't waste human energy on work the agents can do themselves but there will be plenty of work they can't. If it's verifiable, send it off to the agents, if it isn't then bring in the human judgement.

Workflows are the natural solution to the Memento problem

The savant with no short-term memory is the other half of the picture. Highly capable, changes personality to match whatever it's working on, and degrades fast when its context gets polluted. Performance depends on keeping context clean and focused on the specific job in front of it.

My answer is that workflows will stay core to stretching agent capability, and I don't think that changes as models improve. The way you deal with the Memento problem is by breaking work into focused chunks an agent can hold in a single context window, with defined handoffs between them.

That's all a workflow is. Not a project management ceremony, but a way of cutting a problem into pieces small enough to think about clearly. Workflows won't be the only solution to the Memento problem. Context windows will get longer, memory solutions will get better, but even dramatic improvements in both won't remove the fundamental constraint.

The challenge is then how do you build the workflows that make the whole system perform best. Make the steps too small and you're repeating context. Make them too large and performance degrades. Pull in humans too frequently and you're wasting attention, wait too long and they'll have drifted too far off course.

Maximize the value of human attention

That challenge becomes Cyboflow's driving product principle: maximize the value of human attention. Each incremental feature earns its place by making the most of eyeballs looking at it. Some of the ways that manifests:

1. Multi-step, multi-agent workflows with human checkpoints. The core of the product is built around running complex multi-agent, multi-model workflows but critically with designated human steps within them. These are the moments that require judgement and the goal of the system is to minimize those interruptions and maximize progress between those check-ins.
2. Maximize information bandwidth. If you have scarce human attention, you need to make the most of it which means maximizing the useful information you can provide. This is why Cyboflow is not a CLI. CLIs are incredibly powerful but if you want to provide as much information as possible in a way that's easy to consume for humans you need more complex interfaces that text alone can't provide. This is also why Cyboflow is not yet another Slack agent.
3. Minimize context switching, maximize context. Context windows are the fundamental agent constraint, context switching is the fundamental human one. When you are juggling multiple agent streams at once, context switching becomes the ultimate tax that slows the whole system down. Cyboflow is designed to reduce the moments of context switching by providing human context as natively as possible. Session summaries, structured artifacts, design mode, native task tracking, centralized review queue. These features and many others are designed to provide relevant context to users natively and to minimize moments of tabbing between disparate tools.
4. Spin the flywheel faster every turn. Cyboflow isn't just designed to maximize the value of human attention, it's designed to get better at it the longer you use it. Learning loops are at the core of Cyboflow at every level — each run produces findings that improve your codebase, analytics help you measure workflow performance, and A/B testing lets you iteratively improve your workflows over time.

There are many more ways it shows up in the product, but everything comes back to that core driving principle and that's what sets Cyboflow apart. 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.