In a recent AI DevOps podcast, Paul Duvall discussed how agentic AI patterns are powering core engineering disciplines as modern models become more capable. He also shares a repository of agent AI engineering patterns where he documents and advances the practice of AI-assisted software development.
Duvall, author of Continuous Integration: Improving Software Quality and Reducing Risk, positions his pattern collection as an exploration of how established engineering practices are being adapted through the practical use of agent AI in client operations. “Engineering practices become even more important when AI generates code,” he said, emphasizing grounding AI-generated output on shared patterns.
Given the amount of code generated by AI, Duvall emphasized the continued importance of trunk-based development, early and frequent commits, and automated testing, explaining that these will become essential to maintaining quality as change rates increase.
Duvall also described changes in the way developers work with code, stating that “we are currently not reviewing every line of code” when working with AI-generated output, observing that the amount of change is making this increasingly impractical. Instead, Duvall emphasized relying on automated validation and agent guard rails, including codified skills that allow agents to review and refine their own output.
Duvall also discussed how approaches such as specification-driven development are evolving existing engineering methodologies. Duvall’s repository includes example agent-readable specifications for AWS IAM policy generation scenarios that predefine expected behavior, constraints, and acceptance criteria, and enable agents to generate and validate output against well-defined specifications. Describing how the familiar test-first pattern is being adapted to guide AI-assisted workflows, he said:
I’m literally… replicating what I did in Agile and XP… I literally say red, green, refactor… I go through that process.
Duvall also highlighted the challenges early in the agent lifecycle, particularly around defining intent. He noted that while AI tools can generate code quickly, ambiguous or poorly specified inputs often lead to inconsistent or unpredictable results. This has led to a focus on driving agents with clearer specifications, such as structured prompts that explain intent through roles, context, and constraints, leading to specification-driven development and acceptance testing that derives from defined behavior, pointing out that “if you don’t fully explain what your intent is, the results will be random.”
A similar focus on clearer specifications was recently discussed on the DevSecOps Talks podcast with Paul Stack, product director at System Initiative, which works on SWAMP, an agent-based open source platform for infrastructure automation and verification. Stack described how they restructured their development process around agents, going so far as to reject pull requests in favor of Github Issue-based workflows that feed into specification-driven development. he said:
We do not accept pull requests… If you have a design… please open an issue. Explore this interactively and design together.
Appearing on Scott Hanselman’s podcast, Gergely Orosz, author of The Pragmatic Engineer newsletter, discussed open source projects that forego merging pull requests in favor of “remixing,” where the provided PR is rebuilt by an agent according to project standards. Contrasting this with autonomous agents using fully automated “Ralph loops” where sub-agents iteratively refine solutions until requirements are met, Hanselman acknowledged that while architectural and design “sense” is important in critical systems, an “infinitely patient junior engineer” mentality may be better suited to the struggle.
Stack also emphasized the importance of providing precise architectural patterns and practices so that agents can “generate code in a way that is consistent with the codebase,” in addition to defining architectures, constraints, and testing expectations up front. Similar to InfoQ’s report on Boris Cherny’s agent workflows, Stack said Claude’s “Plan Mode” is used to confirm intent before execution, helping avoid “AI horror stories.”
Duvall also pointed to the importance of shifting to the right and extending these feedback loops into production environments. He discussed how observability, telemetry, and even testing in production can be used to shorten feedback cycles and interpret live signals back into the development lifecycle. Looking to the future, he suggested that AI could lead to smaller, more focused teams, and explained that as coordination overhead decreases and automation increases, there will be a shift toward “one pizza teams.”
Similar to earlier changes in engineering, Duvall suggested that quality is increasingly achieved through automation rather than human inspection. he said:
We have built in mechanisms such that code is reviewed, but it may not be reviewed literally every time.
Duvall and Stack emphasized that AI-assisted development requires a combination of shift-left practices and shift-right feedback, and that definition of behavior and operational state become part of the validation process. Duvall also noted the benefits of AI in analyzing production telemetry more broadly to identify patterns and surface issues earlier.
Duvall’s repository is continually updated and defines structured patterns of maturity levels across development, security, and operational scenarios. Patterns include specification-driven development, codified rules and architectural constraints, atomic decomposition with parallel agents, and observable development of workflows with automatic traceability.
Recognizing the move beyond code-centric development, Orosz reflected that engineering identities and practices will move one level beyond the code itself. he said:
I think there’s more than coding that makes us special. And I think we need to cultivate that.
