Agent AI is now a key part of the engineering process, driving huge execution levers and helping us build more code than ever before. However, the tough question I hear more and more from business leaders is: If we’re shipping code faster than ever, why aren’t our products improving at the same rate?
The reason is that writing code has never been a speed limiter. Defining the right requirements, integrating with complex systems, and maintaining software in real-world environments has always been the hard part. The hard part becomes even more difficult when agents flood the organization with large numbers of new codes. Agents compress execution time. They do not suppress uncertainty, accountability, or operational complexity.
As AI-generated code grows in scale, human vision becomes a major new bottleneck, and engineers lose the context they need to catch agent errors. Companies that understand this will be intentional and forward it even creates new roles thanks to AI. Those who don’t will come to a simpler, more devastating conclusion: cut staff and increase AI spending.
Game book
Irreversible structural decisions require caution because technology moves so fast. Enterprise engineering leaders need a thoughtful playbook to navigate the chaos. How to get started:
Phase 1: Finance and risk management
Protect the downside – secure the infrastructure and stop the financial bleeding.
-
Treat management as a primary risk: The pressure to integrate AI is real, but giving teams the freedom to experiment without a centralized structure creates fragmented processes, duplicate work, and runaway costs. Organizations must create shared standards while allowing teams to adapt and explore within defined boundaries. This means treating agent configuration like a production infrastructure—versioning, reviewing, and testing before rolling out instructions and skills.
-
Apply least privilege to non-human actors: Never allow an agent to inherit the full permissions of a human operator. Human engineers are given wide access because they have contextual judgment and are ultimately responsible. Deploying agents with human-level access without careful thought creates an accountability gap in your systems. Make a strict separation between them to read and write/execute to enter and mandate approval gates of people in circulation for disruptive or production-altering actions. As agents transition from bid code to autonomously executed tasks, they should be heavily incorporated into your security model.
-
View your wallet: Protect your overall AI budget by enforcing quotas and rate limits for both engineering and production. The cautionary tale is becoming more common: Uber has capped its AI spending since then Burns the 2026 budget until Apriland according to the unnamed company Axios drew an astonishing $500 million anthropic bill within a month due to fugitive agent loops.
Stage 2: Technical strategy
Build the engine: Choose the right models and measure their success.
-
Be multi-model and multi-seller: No model excels at everything. In order to understand where each one excels, it is important to accurately characterize the behavioral and performance boundaries between models, directing specific tasks to the systems best equipped to handle them. Standardizing on a single vendor or model sacrifices capabilities and introduces a critical point of failure. No organization should accept concentration risk in a core engineering function.
-
Border fee: Treat AI as engineering leverage, not just another SaaS expense. Pay for premium border models that provide the highest quality product and again reduce costly work. Ultimately, the cheapest model isn’t the model with the lowest cost—it’s the model that maximizes efficiency while minimizing your downstream risk.
-
Measure what really matters: Deployments, lines of code, and pull requests have never been good metrics for productivity, and with AI they are actively misleading. Instead, target metrics that add to business outcomes (feature adoption, retention) and engineering sustainability (change in defect rate, avoided defects, code survivability over time). Measure task success and rework time per dollar for AI effectiveness. Token counts are handy for leaderboards, but they can’t tell you if tokens are well spent.
Stage 3: Talent and organization
Realign your human capital to handle the new bottleneck.
-
Move engineers from syntax to systems: As agents handle much of the code generation, human vision and architectural compatibility are the new hurdles. Organizations must intentionally evolve their workforce to transition from syntax writers to systems thinkers and agent-managers. Engineers need the training and mandate to guide agent processes, manage complex cross-system integrations, and maintain a common architectural vision that agents may struggle to maintain.
-
Redefine performance and incentives: Traditional metrics such as story points or sprint velocity can be ineffective when an individual engineer can create the product of an ex-team. Consider realigning your evaluation frameworks to better reward extended business impact, cross-system reliability, and effective agent orchestration. If you want systems thinkers who are more strategic, willing to explore and take risks, and build products sustainably, you need to reward them for a higher level of impact, not product volume.
-
Don’t downsize before your strategy is aligned: If you haven’t integrated agent workflows, measured increased output in production, and reworked your roadmap for faster execution, you don’t really know if your needs and capabilities are aligned. Reducing the number of employees before setting this base is not discipline – it is blindness. The goal is not just small teams, but teams that can cover more strategic territory.
Enterprise adoption of AI requires human flexibility
AI does not replace engineering judgment; is a power multiplier for him. Safely accelerates delivery in well-established systems. Accelerates failure in poorly understood systems. We’re already seeing the results: outages, rising technical debt, and unexpected cost increases from poorly managed adoption. These are operational failures, not theoretical risks.
The mistake organizations are making now is not to adopt AI too slowly – it is to adopt it without understanding where it breaks down.
For the C-suite, understanding these dynamics is no longer optional – it’s a defining factor in how business moves in this era. The challenge is that the speed of implementation is outstripping the industry’s ability to manage results. We have introduced the best power tool to engineering teams. The old adage says measure twice and cut once. Instead, many firms simply choose to downsize.
Joe Bertolami is the CTO and co-founder Clifton AI.
