AWS re:Invent 2025 - AI is Breaking the SDLC: Here's How to Fix It (DVT103)

Summary of "AWS re:Invent 2025 - AI is Breaking the SDLC: Here's How to Fix It"

The Rise of Agentic Workflows

• The presentation discusses the growing trend of "agentic workflows" - the use of AI-powered agents to automate various software development tasks.
• In the past, developers would manually copy-paste between different tools (terminal, IDE, ChatGPT, etc.). Now, more advanced IDEs and "headless agents" are emerging that can perform a wider range of automated tasks.
• Analysis of GitHub activity shows a rapid growth in the usage of these AI agents, with a shift from primarily doing code review and issue triage to actually pushing code directly to repositories.
• This trend is also observed in the speaker's company, CircleCI, where they see a significant increase in build activity driven by these AI agents, indicating they are being used for economically valuable work, not just experiments.

The Problem: Delivery Processes Can't Keep Up

• While the ability to generate code faster is beneficial, the speaker argues that the downstream delivery processes have not kept up, leading to several issues:

  • Massive pull requests that are difficult to review
  • Increased build instability and delays
  • Minimal impact on actual product effectiveness
  • Increased developer burnout

• Using queuing theory, the speaker demonstrates how the mismatch between the speed of code generation and the speed of delivery can lead to growing backlogs and delays, even if the individual components are becoming more efficient.

• Industry data suggests that while the top performers are seeing around 10% improvements from AI, average and below-average teams are actually experiencing negative effects.

The Solution: Focus on Validation and Context

• The key to addressing this problem is to focus on improving the validation and delivery processes, not just the code generation capabilities.
• Validation is the critical step of checking the quality of the work before it is merged or deployed. This is a scalable, durable, and tractable approach for most organizations, compared to trying to build and train their own AI models.
• Providing the right context to the AI agents is also important. This includes:
  • Feeding the results of validation pipelines back into the agents to improve their performance
  • Controlling the tools, tasks, and prompts given to the agents
  • Taking a bottom-up approach to build a library of validated, reusable agent capabilities

CircleCI's Approach: Chunk

• The speaker introduces CircleCI's new agent, Chunk, which is designed with these principles in mind:
  • Automatically executing CI pipelines to validate any changes made by the agent
  • Focusing on improving the reliability and speed of the delivery process, such as by addressing flaky tests and build optimization
  • Leveraging the context of successful and failed changes to continuously tune the agent's performance

Key Takeaways

• The adoption of AI-powered agents in software development is a real and growing trend, with agents now responsible for writing and pushing code, not just code review and issue triage.
• This trend is causing issues for many organizations, as their delivery processes cannot keep up with the increased speed of code generation.
• The solution lies in focusing on improving the validation and delivery processes, rather than just the code generation capabilities.
• Providing the right context to the AI agents, such as feedback from validation pipelines, is crucial for ensuring their reliable and effective performance.
• CircleCI's Chunk agent is an example of a tool designed with these principles in mind, aiming to make the delivery process faster, more reliable, and more effective.