Coding Agents Won't Scale on Their Own; Neither Will Teams

Patrick Debois draws a parallel between the current skepticism surrounding AI agent adoption and the resistance faced by continuous delivery in 2009. He recalls being told that continuous delivery was a "crazy idea" back then, and observes a similar sentiment now regarding the "dark factory" concept. This suggests that new technological paradigms often encounter initial pushback, not because of technical infeasibility, but due to organizational unpreparedness to embrace change.

Debois explains that the common objection, "it will not work here," often masks the underlying issue that organizations are simply "not ready yet" to adopt new methods. He asserts that technology is rarely the primary barrier to innovation; rather, it is the organizational readiness and willingness to adapt processes and mindsets that determine successful integration.

Patrick Debois predicts that the technical implementation of agent loops, including their optimization, will eventually become a commodity service. He suggests that frontier labs and specialized companies will offer these capabilities, making them readily available rather than requiring individual organizations to develop them from scratch.

According to Debois, this commoditization means that the true differentiator for companies adopting AI agents will not be how they build the agent loop itself. Instead, success will hinge on how effectively organizations integrate these agents into their existing workflows, optimize team interactions, and foster systemic improvements around the agents.

Debois emphasizes that Conway's Law continues to be highly relevant in the context of AI agents. This principle states that organizations design systems that mirror their own communication structures, implying a deep link between organizational structure, team communication, and the tools they use.

In the AI era, this means that the focus must shift beyond merely improving individual agents. Instead, organizations need to consider how teams interact with these agents, how platforms facilitate this interaction, and the broader organizational changes required to maximize their effectiveness.

Patrick Debois acknowledges the prevailing narrative that developers are evolving into conductors and orchestrators of AI agents. This shift reflects a broader evolution in development roles, where engineers manage and guide AI rather than solely writing code from scratch.

However, Debois notes that this transition initially creates friction, as many engineers did not sign up for roles that primarily involve "prompting" or simply guiding AI. There is a perceived loss of direct coding involvement, which can lead to dissatisfaction among developers accustomed to hands-on creation.

Debois identifies a new technical path that can restore engineers' sense of identity and engagement when working with AI agents. By introducing harnesses, loops, and more autonomous work within the organization, a fresh frontier for engineering opens up.

This approach allows developers to programmatically assist agents by building specialized tooling, shifting their role from simple prompt engineering to more complex systems engineering. When engineers can directly contribute to improving how agents operate through code and infrastructure, they feel more engaged and their technical skills are better utilized.

Patrick Debois proposes that the skepticism and even "anger" of developers regarding the quality of agent-produced output can be a valuable asset. Instead of dismissing resistant individuals, organizations should channel their complaints into constructive improvements for AI agents.

By empowering skeptical developers to refine agent context and build better harnesses, their critical eye can directly contribute to enhancing the agents' performance and reliability. This approach transforms resistance into a mechanism for system improvement, leading to more robust and higher-quality AI outputs.

Patrick Debois advises companies to implement a fundamental mentality shift among their developers: stop merely fixing the code produced by AI agents, and instead focus on improving the underlying system. He argues that patching individual code outputs is a reactive and inefficient approach that fails to address the root causes of issues.

By improving the system that agents operate within—including their harnesses, context, and feedback loops—organizations can prevent recurring problems and enhance the quality of agent-generated code for the entire team. This proactive approach leads to more scalable and sustainable productivity gains across the development lifecycle.