Augment Code 调研 219 位工程负责人：AI 原生开发中的信任与角色落差

AI changed software development. Nobody agrees on the rules now.

We surveyed 219 engineering leaders going AI-native. Half their code is now AI-generated, and they're not sure they trust it.

Three months ago, we pushed hard on agent-generated code at Augment. PR volume went up. Confidence in what was shipping went down. We took two days off the roadmap to ask the team how they were feeling. The word cloud came back with frustration and uncertainty right alongside curiosity and excitement. One group named the shift directly: from "proud builder" to "proud coordinator."

Then we started talking to other engineering leaders and realized we weren't the only ones. The same stuff kept coming up. Nobody knows how to review code that agents wrote. Engineers are asking their managers whether their skills still matter, and the managers don't have an answer. People are shipping faster than ever and trusting what they ship less than ever.

We wanted to know if this was a pattern or just our circle. So we ran a survey.

219 engineering leaders responded, from managers to CTOs, running teams that range from 1 to 1,000+ engineers. The sample skews AI-forward. Here's what they told us.

• 48% of their code is now AI-generated.
• 55% are concerned or very concerned about losing shared understanding of the codebase.
• 63% say their engineers are raising fears about skill relevance to their managers.
• At the largest teams (201-1,000 engineers), that last number rises to 89%.

When we asked for three words to describe how they feel about AI and the future of software engineering, many put contradictory emotions in the same answer: "excited, anxious, invigorated." "Insecurity, mistrust, hope." "Optimistic, excited, threatened."

We'd expected two groups: the optimists and the worried. Instead, it's the same people, holding both feelings at once.

The full report maps where the strain is showing: technical debt in the code, cognitive debt in the people, and intent debt in the specs, role definitions, and onboarding processes that haven't been updated. It walks through five contradictions in the data. The sharpest one: leaders are hiring for a different job (the #1 hiring priority is now evaluating AI-generated code; traditional coding has dropped to #5), but only 19 of 219 orgs have formally changed their role definitions.

Every problem in the report traces back to the same gap: individual engineers have adopted agents, but the organization hasn't changed around them. That's what we're building Cosmos to address: an operating system for agentic software development where humans, agents, and institutional knowledge are coordinated at the org level.