Understanding how AI impacts each stage of development and identifying where it delivers the most value within the lifecycle is key to integrating these capabilities at scale within an organization. All of this must be approached while maintaining the flexibility required in an ecosystem of tools that evolves almost weekly, with platforms quickly rising and falling in relevance.
A common mistake is to assume that success with agents depends solely on the underlying LLM. In reality, the key lies in what we call the agent harness: the combination of tools, methodologies, and context that empowers the model. On its own, the model is not enough.
To achieve the desired outcomes, at Baufest we guide agents just as we would a junior developer. This involves feeding them with documentation, rules, code examples, and the specific skills of our teams. We also implement feedforward and feedback practices to create a positive learning loop, enabling the agent to adopt our methods and intentions and align precisely with each project’s processes.
Within this methodological framework, it is essential to provide agents with the tools needed to validate their work autonomously—such as unit tests, static analysis, linting, and code reviews. This allows agents to iterate independently until they meet the project’s quality standards.
By shifting these validations earlier in the software development lifecycle—before integration or commit—we can detect errors faster and at a lower cost. Today, the real bottleneck is no longer code generation, which trained agents already handle effectively, but verification. Automating quality controls is the only viable strategy to prevent senior engineers from being overwhelmed by endless manual reviews.
Of course, governing these technologies goes beyond simply assigning licenses. It is crucial to understand how each tool will be used, evaluating both its impact on productivity and its effect on the economic equation, especially as AI costs continue to rise. This requires a clear understanding of each model’s strengths to determine when to use them. For example, more powerful and costly models—such as Opus—can be reserved for high-complexity tasks like initial planning, critical bug resolution, or deep logical reasoning. In contrast, for rapid iterations, minor refactoring, or tasks where prior context already guides the AI effectively, a more cost-efficient model like Sonnet is more than sufficient.
Ultimately, to achieve real results, agentic AI must be applied across the entire software development lifecycle (SDLC), not just during the build phase. This includes everything from problem definition and requirements to the updating and maintenance of legacy systems. It is also important to rely on reliable metrics to measure team productivity end to end, rather than relying solely on perception.
Even if developers feel more productive when using agents to generate code, if review and deployment processes are not optimized to handle the volume of AI-generated output, a new bottleneck will emerge. In this new landscape, the role of the senior engineer evolves: beyond programming, they must also act as a trainer, architect, and auditor of AI, leading a virtuous cycle of productivity where quality control becomes a constant driver of improvement.
