The market has made its bet.
Coding agents are no longer being priced like nice-to-have tools. They are being priced like infrastructure. In May, Blitzy announced a $200 million round at a $1.4 billion valuation. In April, TechCrunch reported that SpaceX had secured an option to buy Cursor for $60 billion, after Cursor's November round valued it at $29.3 billion.
This is a control story, not an autocomplete story.
The easy reading is that the engine keeps getting better. Agents write more code. They touch larger codebases. They can sit inside enterprise workflows instead of hovering around the edge. That part is true. It is also the least interesting part.
The harder problem is what happens when you have more than one agent, more than one task, and more than one human trying to keep up.
Brian Chesky described the interface problem cleanly in Airbnb's May AI discussion. He said chatbots are the wrong interface for travel and e-commerce because they have four problems: too much text, no direct manipulation, poor comparison, and no multiplayer collaboration. He was talking about travel, but the diagnosis lands harder in software.
Too Much Text
Chat works when one model answers one question.
It breaks when an operator is staring at five agents, three branches, two failing tests, and a stream of tool calls that all need judgment. The human is not reading a conversation anymore. They are doing traffic control.
That is why so many agent products feel useful in demos and exhausting in real use. The text is the interface. The text is also the problem.
No Direct Manipulation
Most agent systems still make the user steer by describing what they want in language. That is backwards.
If I need to pause an agent, narrow its scope, reroute it to another file, compare two candidate patches, or promote one result and reject another, I do not want to write a paragraph about it. I want handles. I want obvious controls. I want to move things around directly.
Sliders beat speeches when the task is operational.
Poor Comparison
This is the quiet killer.
A lot of knowledge work is comparative. You do not want one answer. You want three candidate implementations, two summaries of a failing test, or side-by-side plans with different tradeoffs. Chat turns all of that into a single stream.
Fine for conversation. Terrible for judgment.
If a product cannot show alternatives cleanly, it forces the human to reconstruct the comparison in their head. That adds friction exactly where the human is supposed to add value.
No Multiplayer
Real work is social.
An engineer, a manager, a designer, and a reviewer may all need to inspect the same agent output, leave comments, and make a decision from the same state. Most agent interfaces are still single-player terminals with nicer branding.
A solo prototype can live with that. Production work inside an organization cannot.
The model can be useful. The interface can still be wrong.
That is the part the market keeps underpricing.
Airbnb is a good example of the split. The company says roughly 60% of new code is AI-written and that its customer support bot handles about 40% of issues without escalating to a human. So the model layer is already doing real work. Chesky's point is sharper than "chatbots are bad." The useful model still needs a richer operating surface around it.
Coding agents are in the same place.
Blitzy's pitch makes that obvious. The product goes beyond code generation: persistent codebase understanding, long-running inference, and orchestration across thousands of agents in parallel. Once you are talking about that kind of system, you have left the model story behind. You are talking about a control plane.
And control planes create their own moat.
The model layer is becoming easier to rent. Every serious product can call Claude, GPT, Gemini, or some mixture of them. The edge is moving upward. If every company can access the same frontier models, "we have an agent" stops being much of an advantage. "We built the interface humans actually want to use" starts looking like the real one.
That interface will probably feel more like an operating surface for work than a chatbot.
It will have direct manipulation instead of prompt-only steering. It will show candidate outputs side by side. It will make delegation visible. It will make uncertainty visible. It will let a team share state instead of turning everything into private transcripts. It will make rollback easy.
Harder to build. More defensible when it works.
The headline money is chasing the engine because the engine is legible. You can benchmark it. You can demo it. You can tell a clean investor story around it.
The steering wheel is messier. It is slower to build. It is less glamorous. It is also where the real friction lives once agents stop being toys and start being part of how teams work.
If agent capability keeps improving, the bottleneck moves from generation to supervision. Once that happens, the company that makes agents easy to steer will matter more than the company that makes them slightly better at typing.
That is the moat.
The next big winner in coding agents will probably be defined by control, not flash: the product that makes a human team feel in charge of a small swarm without reading every line of every transcript.
The market is funding the engine. The harder business is the steering wheel.