AI Agents: The Rebuilt CI/CD Pitfalls

Sumaiya Shrabony, Technical Program Manager at the University of Colorado Denver, presents a compelling argument in her talk: "Every Solo Agent Builder Eventually Reinvents a Worse Version of CI/CD." Shrabony contends that while agent demos often present a smooth, successful 'happy path,' the reality of agent operations is far more complex, riddled with potential failures that are often overlooked. She highlights that the true challenge lies not in the agent's output itself, but in the underlying processes and the lack of robust verification, drawing parallels to the evolution of Continuous Integration/Continuous Deployment (CI/CD) in software development.

The Illusion of the Happy Path

Shrabony begins by illustrating the common perception of agent development, where initial demos showcase a seamless flow from input to output. However, she emphasizes that this 'happy path' is often incomplete and misleading. The real work begins when the agent's initial promise falters, leading developers to build custom solutions for issues that have already been addressed by established software engineering practices.

She outlines a typical agent workflow, which includes stages like scheduling, command execution, research, planning, production skill execution, verification, and final output. Each of these 'handoffs' is a potential point of failure, where the system can subtly deviate from expected behavior. Shrabony notes that when building agents solo, developers often find themselves recreating processes that already exist in more mature forms, such as CI/CD pipelines.

Five Common Failures in Agent Systems

Shrabony details five specific failures that commonly emerge as developers build their own agent systems:

• Voice Drift: The content generated by the agent starts to deviate from the intended voice or tone, often defaulting to generic marketing language. This happens because the system lacks a mechanism to enforce the desired voice patterns.
• Missing Verification: Claims made by the agent, such as a reduction in rework by a certain percentage, are presented without any underlying verification or source trail. This means the "evidence" for the claim is absent, making it untrustworthy.
• Duplicate Hook: The agent's opening statement or 'hook' becomes repetitive or recycled, which can erode audience trust and signal a lack of originality.
• Staging Gates: Artifacts are produced and appear 'ready' but lack essential checks or validation before being shipped. This is akin to shipping code without running tests, leading to potential downstream issues.
• Rollback & Audit: When something goes wrong, the system lacks the ability to trace the exact steps, skills, or handoffs that led to the failure, making debugging and reconstruction difficult.

Shrabony illustrates these failures with concrete examples from her own experience and showcases how implementing specific 'gates' or checkpoints can prevent them. For instance, in the case of 'Voice Drift,' running the agent in 'Guarded Mode' with a voice contract prevents the output from deviating from the desired patterns.

The Importance of Boundaries Over Platforms

Shrabony argues that the solution is not to rely on complex platforms or frameworks, but rather to implement essential 'boundaries' or gates within the agent system. These boundaries act as crucial checkpoints to ensure the quality, integrity, and reliability of the agent's output.

She identifies five key gates that developers should consider:

• Output Contract: Ensures the artifact meets the required shape and format before being saved.
• Voice Contract: Verifies that the output matches the system's intended voice and style.
• Verification Contract: Checks that all claims made by the agent are traceable to a verifiable source.
• Dedup Check: Determines if the output is genuinely new or if the system is recycling previous content or angles.
• Audit Trail: Allows for the reconstruction of events when failures occur, making debugging and troubleshooting possible.

Shrabony concludes with a powerful takeaway: "In software, we learned not to deploy just because code exists. In agent systems, we need to learn not to ship just because the artifact looks complete." She urges developers to map their agent's handoffs and identify the most expensive or critical handoffs to implement these crucial verification gates, thereby building more reliable and trustworthy AI agents.