Latent Space sees the AI race moving from model scores to agent work

Latent Space’s latest AINews edition captures a shift the industry has already been showing through products: leading model labs are no longer acting only like factories for new weights and benchmark scores, but like agent labs. The original piece, “All Model Labs are now Agent Labs”, is not built around a single loud launch. The more important move is quieter: the model has become a component, not the whole product.

That distinction matters. A classic model lab optimizes a system’s ability to answer a prompt, write code, summarize a document or pass a test. An agent lab has to solve a more stubborn problem: how the model receives a goal, plans steps, chooses a tool, checks the result, remembers context and knows when to stop. That is why models are increasingly surrounded by agent frameworks, computer control, browsers, connectors, sandboxes and evaluations that measure execution over time rather than a single response.

OpenAI has formalized that direction through the Agents SDK, where the model sits inside a broader system with tools, execution traces and workflow control. Anthropic pushed the same idea toward interfaces not originally designed for models with computer use. Google has tied search, data and enterprise workflows into agent logic through Vertex AI Agent Builder and related products. The implementations differ, but the direction is the same: capability is moving out of the isolated prompt and into the operating system around the model.

For the TECH&SPACE audience, the cold takeaway is simple: agents are not just “chatbots with buttons.” If a system can read documents, call APIs, edit files, run code or control a browser, the product is no longer selling only language fluency. It is selling trust in execution. That raises new questions about permissions, audit trails, test environments, rollback, data privacy and responsibility when an agent takes the wrong step.

Evaluations have to change as well. A model that shines on a static test can fail inside a long task with ambiguous intermediate decisions. An agent that plans well can burn too many calls, select the wrong tool or confidently finish a task it never verified. The useful metric is not only whether the final answer is correct, but how stable the system is, how often it recovers from mistakes and whether a human can understand why it took a given action.

This shift does not mean models have stopped mattering. Better models give agents a wider operating range. But the next difference between labs will not be visible only in who has the stronger base model. It will show up in who can build an agent layer precise enough for real work, transparent enough for oversight and constrained enough to avoid becoming expensive automation for bad decisions. Latent Space’s headline lands because it names the moment cleanly: the model lab is now the starting point, while the agent lab is where AI becomes either a usable product or just another impressive demo.