Look at where the money went this cycle:
- Cursor (Anysphere): reportedly $3.4B raised at a $29B valuation, with reports of a SpaceX-linked option at $60B
- Replit: roughly $870M raised at a $9B valuation
- Lovable: about $550M raised at a $6.6B valuation
- Blitzy: $200M at $1.4B for parallel coding agents
Investors are treating coding agents as a foundational layer, a bet that agents become something the rest of the software economy sits on, the way cloud compute did. This essay argues the bet is aimed at the wrong layer. The engine is getting funded. The steering wheel, the interface where humans actually direct and supervise agents, is where the constraint now lives.
The evidence that agents already work
Airbnb is not a coding-agent startup, which is exactly why its numbers are useful. It says around 60% of its new code is AI-written, and its customer support bot handles roughly 40% of issues without escalating to a human. Those are production numbers at a large consumer company. Meanwhile Blitzy raised specifically to run coding agents in parallel, and every frontier lab now sells agentic coding, though much of what makes it usable in practice comes from scaffolding built around the model rather than the raw model itself.
Frontier models are still meaningfully differentiated on coding ability. Benchmark gaps are real, and anyone choosing a model for serious work feels them. But the direction of travel matters. As more models clear the bar for writing code, browsing, and calling tools, raw capability starts looking less like an edge and more like an input. If that convergence continues, the question becomes what differentiates the products built on top. My answer: the interface.
Where chat breaks
Airbnb's CEO named the bottleneck in the context of consumer products, but his diagnosis of chat interfaces generalizes to developer tools and enterprise work. It came down to four problems:
- Too much text. Everything gets flattened into a scrolling transcript.
- No direct manipulation. You describe changes instead of making them.
- Poor comparison. Evaluating three options side by side is painful in a linear thread.
- Not multiplayer. Chat assumes one human and one agent, but real work involves teams.
Chat works fine for the tasks agents were first sold on: generating code, drafting text, transforming structured data. Those are linear, single-player, text-native tasks. It degrades fast for anything visual, spatial, comparison-heavy, or collaborative, which describes most complex knowledge work.
Ask an agent to refactor a function and chat is a good fit. Ask it to help your team evaluate three architecture proposals against cost, latency, and migration risk, with two engineers and a PM weighing in, and you are fighting the interface the whole way. You end up copying outputs into docs and spreadsheets, which is a signal the interface failed.
The multiplayer problem gets worse with more agents
Parallel agent deployment makes this sharper. Once you have several agents working simultaneously, a human supervisor needs to compare their outputs, intervene mid-task, and redirect effort. A chat transcript per agent does not scale to that. You need something closer to an air traffic control view than a messaging app.
Anyone who has actually run multiple agents in parallel hits this wall quickly: the limiting factor is coordination and supervision UX, not model capability or compute.
What follows if this is right
The core claim of this piece: if capability converges while the interface problem stays unsolved, the lasting advantage sits at the interaction layer rather than the agent layer. Three things follow:
- Current coding agent valuations assume the agent layer captures the value. If interfaces are the bottleneck, a meaningful share of that value migrates to whoever solves human-agent collaboration for complex domains.
- Multiplayer AI interfaces, meaning shared workspaces where teams and agents operate on the same artifacts, are underfunded relative to the engine layer.
- Regulated and high-stakes sectors will force the issue first, because human oversight there is mandatory.
What to watch
- Adoption metrics for agentic coding tools, and specifically whether usage plateaus at single-player tasks
- Whether coding benchmark gaps between frontier models narrow or persist, since persistent gaps weaken the convergence premise
- Startups shipping multiplayer or direct-manipulation interfaces for agent supervision
- Airbnb's AI search experiments, as a test of moving past chat in a consumer product
- Whether parallel agent platforms like Blitzy win in regulated sectors, where supervision UX is mandatory
The engine got funded. The steering wheel is still open. That is where the interesting work is.