Organizational Structure for AI-First in the Harness Era

This episode of 'Silicon Valley 101' invited three founders of Creao to discuss Harness Engineering. It's not just simple Prompt Engineering, nor Context Engineering, but truly treating AI as the primary productive force and redesigning the entire company organization and workflow.

Over the past three years, LLM engineering capabilities have gone through three evolutions.

In 2023, people were talking about Prompt Engineering, focusing on how to write good prompts; in 2024, it shifted to Context Engineering, focusing on providing more complete context to models; by 2026, a new term suddenly emerged: Harness Engineering.

The word "harness" originally means the reins for controlling a horse, but now it describes how to build a system around LLMs that can self-heal, self-improve, and operate continuously in the real world.

Creao's CTO Peter previously worked on Meta's Llama team and Apple's multimodal group. He posted a tweet that exploded to 1.87 million views.

The core idea: most companies' so-called "AI-First" is fake. Real AI-First means transforming AI from an auxiliary tool into the productivity driver, completely re-architecting the organizational structure and workflow.

The result: they can finish a feature at 10 a.m., run an AB test at noon, kill it based on data by 3 p.m., and rewrite a better version by 5 p.m.

Traditional processes might take six weeks; now it's done in a day.

In the past, product development was slow, allowing marketing teams to prepare four to five months in advance. Now it's reversed: development speed far exceeds marketing, and marketing teams are chasing the features engineers release.

Clark says they no longer need a feature wishlist or bug list. Because agents can detect and auto-fix bugs in time, and features are so abundant they can't even use them all.

This change is dramatic.

What used to be the scarcest resource in a company—development capacity—has now become "whether the organization can absorb AI capacity."

Trust has shifted from "trusting people" to "trusting AI."

Another big change: the way cross-team collaboration works has transformed.

Before, cross-team alignment costs were extremely high, with constant back-and-forth syncs among engineering, product, and marketing. Now AI can directly tell Marketing: here's what Engineering is releasing today, without humans acting as go-betweens.

Peter and his team even dismantled the product manager role.

Because from their perspective, alignment costs are themselves a false need. If the system is powerful enough, decisions can be driven by AI, and information can flow automatically via AI.

Even more counterintuitive: junior engineers adapt more easily to an AI-First environment than senior engineers.

Senior engineers carry too much tech debt and mental inertia, unwilling to expand their scope into product design and data analysis. Junior engineers, on the other hand, are more willing to cross functional boundaries.

Peter says that the most valuable people in the future may not be deep specialists in a vertical domain, but generalists who possess Architecture + Product Sense + Marketing Sense.

This is worth pondering.

Once AI amplifies execution ability, what becomes truly scarce is: can you define problems, judge direction, understand users, read data, and build the system?

What really blew my mind was Clark's mention of the Agent Economy.

They found that a lot of marketing materials today are actually optimized for agents, not humans.

In the future, decisions like buying things, subscribing to newspapers, or ordering milk might all be made by agents. So, is the content you produce today meant for human consumption or AI consumption?

The criteria for value judgment may become completely different.

In the past, we optimized for human attention; in the future, we may also need to optimize for agent understanding, retrieval, judgment, and action.

They divide Harness into two layers.

The first layer is the harness for the company's internal agent systems, focusing on self-healing and auto-fixing.

The second layer is the agents that users build themselves on the Creo platform, which can also continuously self-improve.

Their entire team is only 25 people, with fewer than 10 engineers, yet they completed the entire architecture overhaul in two weeks.

Peter says that a product of this scale a year ago would have required at least 100 people working for four to five months.

This isn't to say that the AI-First transformation comes without cost.

They spent a long time aligning the entire team's mindset, especially getting everyone to truly believe: AI can lead, not just assist.

Many people say they embrace AI, but essentially they are only willing to use AI tools to boost their own efficiency, not to cede the role of primary productive force to AI.

This mental gap determines whether a company can truly complete its AI-First transformation.

After listening to this episode, my biggest cognitive upgrade is:

AI-First is not about layering AI tools on top of existing processes, but treating AI as a new species and redesigning the entire organizational ecosystem.

The core value of humans in the future will not be execution, nor alignment, but defining the direction of needs and ultimately reviewing whether outcomes serve human interests.

This episode is incredibly dense with information; the full transcript is very long, so I recommend going to Podwise to read the complete version.

Do you think the product manager role will disappear in the future, or will it evolve into a new form?

Feel free to discuss in the comments.

More hardcore podcast content: Podwise @PodwiseHQ