In a recent LinkedIn post, Luís Rodrigues discusses the significant risks associated with relying solely on autonomous AI agents for software development, arguing that the current hype around rapid AI-driven app creation can be misleading. Rodrigues, a proponent of structured development, cautions against the notion of “vibe coding,” where speed is prioritized over robust engineering principles.
He highlights a personal experience using an AI tool over the holidays, which enabled him to build a LinkedIn efficiency tool rapidly. “The speed was real. Days, not weeks,” Rodrigues shared. However, he contrasts this efficient, human-steered process with the current trend of handing over entire development tasks to autonomous agents, a practice he believes is fraught with peril.
“One agent wrote the plan. The same agent wrote the code. And the same agent graded the work.”
Rodrigues elaborates on the potential for errors and lack of accountability when a single AI agent manages the entire development lifecycle. He points out that the initial specifications can become lost or corrupted through numerous iterations, leading to fundamentally broken features despite passing initial tests. “The spec from turn 3 is gone by turn 90. Retries run forever because nothing counts them,” he states, emphasizing the systemic issues with such an approach.
The Importance of Structure in AI-Assisted Development
Rodrigues advocates for a more structured approach to AI in coding, likening the current autonomous agent model to a single contractor who also performs their own inspections. He proposes a system built on discrete, specialized components, which he metaphorically describes as a graph or a crew with a foreman.
Key Principles for Robust AI Development
In his post, Rodrigues outlines several critical principles for building reliable systems with AI agents:
- Nodes with one job each: Each component should have a single, clearly defined responsibility.
- Tools scoped per node: AI tools should be limited to the specific task of their assigned node.
- Artifacts passed, not chat history: The system should rely on passing discrete outputs (artifacts) rather than extensive conversational logs, ensuring clarity and traceability.
- Verification by an agent that reads the spec: An independent AI agent should be tasked with verifying the output against the original requirements.
- Bounded retries: Mechanisms should be in place to limit and manage retries to prevent endless loops.
“The fix it’s structure,” Rodrigues asserts, underscoring that the solution lies in architectural design rather than simply increasing the power of individual agents.
The Craftsman and the Tool
Rodrigues concludes by emphasizing that while AI agents are incredibly powerful tools, they still require human oversight and skill. He draws an analogy between an autonomous agent acting as both developer and inspector and a construction crew with a foreman overseeing the work. “Agents are the most powerful tool we’ve ever had. But a tool still needs a craftsman,” he argues.
This perspective suggests that the future of efficient and reliable software development lies not in full automation by single agents, but in a synergistic relationship where human expertise guides and validates the capabilities of AI tools. Rodrigues is set to teach this methodology further in an upcoming workshop.
📝 About This Content
This article is based on insights shared by Luís Rodrigues on LinkedIn.
📅 Originally posted on September 7, 2026 | View original post on LinkedIn →