Coding agents are only as good as the signals you feed them

The New Stack

by Arjun Iyer

February 23, 2026

AI-Generated Deep Dive Summary

The effectiveness of AI coding agents is heavily dependent on the feedback loops and infrastructure they operate within, according to recent insights from the field of DevOps and cloud engineering. While advancements in code generation have made agents more capable, most teams still rely on manual workflows that bottleneck productivity. The key issue lies in the validation process: agents generate code, but humans ultimately test it, deploy it, and address integration issues, limiting the potential efficiency gains AI could deliver. Organizations like Stripe, Ramp, OpenAI, and Anthropic have demonstrated that investing in robust feedback infrastructure can unlock significant productivity improvements. These companies have built systems where agents are equipped with tools and constraints to operate effectively, enabling them to verify their own work and iterate without human intervention. For example, OpenAI’s “humans steer, agents execute” approach involves writing docstrings and assertions for the model, allowing it to generate code that is automatically tested and refined until it meets specified criteria. Similarly, Stripe’s Minions framework leverages an MCP server called Toolshed, which provides over 400 tools and integrates deterministic verification steps into the agent’s workflow. When a Minion generates code, it undergoes a rigorous testing process, including git operations, linting, and formatting checks. This approach ensures that agents can operate with greater autonomy, reducing reliance on manual oversight. For DevOps professionals, this shift toward harness engineering represents a game-changer in how AI is integrated into software development. By focusing on building robust environments and feedback loops for agents, teams can elevate their productivity while maintaining control over the outcomes. This approach not only accelerates development but also ensures that AI models are aligned with specific business needs and constraints, ultimately transforming engineers into architects who guide rather than execute every task. The lessons from these early adopters highlight the importance of rethinking traditional workflows to fully harness the potential of AI in coding. By prioritizing infrastructure over code generation capabilities alone, organizations can unlock the speed and scalability that AI agents promise, making them true partners in the software development process.

Verticals

devopscloud

Originally published on The New Stack on 2/23/2026