The CEO of one of the most prominent AI coding startups is pushing back against the idea that his own technology should replace software engineers. Scott Wu, who leads Cognition — the company behind Devin, an AI agent that can write and debug code autonomously — recently argued that the real value lies in human-AI collaboration, not wholesale automation.
His position matters because it comes from inside the industry building these tools. If the people creating AI coding agents are cautioning against full automation, engineering leaders should pay attention.
Why the “Replace Everyone” Approach Fails
The temptation is obvious. AI coding agents can generate working code in minutes, handle repetitive tasks, and work around the clock. Some companies have already experimented with dramatically reducing engineering headcount.
But the early results are concerning. Teams that lean too heavily on AI-generated code report mounting technical debt — the accumulated cost of shortcuts that make future changes harder. Code review backlogs grow because someone still needs to verify what the agent produced. Knowledge transfer breaks down when fewer humans understand how systems actually work.
Wu’s argument is essentially practical: AI agents excel at execution but lack the judgment to make architectural decisions, understand business context, or mentor junior developers. Treating them as replacements rather than tools creates problems that compound over time.
The Governance Framework Your Teams Need
For CIOs and engineering leaders, the question is not whether to use AI coding agents — that decision is largely made. The question is how to structure governance so you capture productivity gains without creating new risks.
Start with clear boundaries. Define which tasks agents can handle autonomously (boilerplate code, unit tests, documentation) versus which require human initiation and review (anything touching security, core business logic, or external integrations). Document these boundaries and make them part of your engineering standards.
Next, redesign code review. Traditional review assumes a human wrote the code and understands it. AI-generated code needs different scrutiny — reviewers should verify not just correctness but also maintainability, since agents often produce functional code that is difficult for humans to modify later. Some teams now require authors to annotate AI-generated sections explicitly.
Finally, invest in observability. Track what percentage of your codebase comes from AI agents, how often that code requires rework, and how it performs in production. Without this data, you cannot make informed decisions about expanding or restricting agent usage.
Rethinking Performance Metrics
If your engineers are now producing more code with AI assistance, your old productivity metrics are broken. Lines of code per developer was always a flawed measure; now it is meaningless.
Forward-thinking engineering leaders are shifting toward outcome-based metrics: time to ship features, defect rates in production, and how quickly teams can respond to incidents. These measures capture whether AI tools are actually helping or just creating the illusion of speed.
Individual performance evaluation needs adjustment too. Engineers who effectively direct AI agents, catch their mistakes, and maintain system-level understanding are now more valuable than those who simply write code quickly. Update your rubrics to reflect this before your next review cycle.
The Upskilling Investment You Cannot Skip
Here is the uncomfortable truth: AI coding agents raise the floor for what engineers can produce while also raising the ceiling for what organizations expect. Your team needs new skills to stay effective.
The critical capabilities are prompt engineering (how to give agents clear instructions), architectural thinking (understanding systems well enough to evaluate agent output), and integration expertise (connecting AI-generated components with existing infrastructure). None of these are automatic, and none come from simply giving engineers access to new tools.
Budget for training now. The companies that treat AI coding agents as a free productivity boost without investing in their people will discover the hidden costs — in bugs, in rework, and in attrition from frustrated engineers who feel set up to fail.
What This Means for You
The debate over AI coding agents is settling into a practical consensus: these tools work best as powerful assistants, not autonomous replacements. Wu’s comments from Cognition reinforce what many engineering leaders are learning through experience.
Your action items are concrete. Audit your current agent usage and establish formal governance boundaries. Update code review processes to handle AI-generated output. Shift performance metrics toward outcomes rather than output volume. Allocate upskilling budget before your next planning cycle.
The organizations that get this right will ship faster with fewer defects. Those that chase headcount reduction without investing in governance will spend the next two years cleaning up the mess. Choose wisely.