Why AI Agents Are ‘Just Another Backend’

Progressive Delivery in the Age of AI: A Discussion with Edith Harbaugh of LaunchDarkly

Key Concepts:

• Progressive Delivery: A software release strategy focused on frequent, small releases with continuous monitoring and control, contrasting with traditional quarterly releases.
• Feature Flags: Techniques to enable or disable functionality without deploying new code, enabling controlled rollouts and experimentation.
• AI Native Approaches: Utilizing AI tools and models directly within the software development and deployment lifecycle.
• Cloud Native Approaches: Building and running applications designed to leverage the scalability and flexibility of cloud computing platforms.
• Guarded Releases: A feature flag implementation that automatically rolls back features based on pre-defined metrics and thresholds.
• AI Configs: LaunchDarkly’s product allowing testing of different LLMs against business metrics in a controlled environment.
• Messy Middle: The current transitional phase in software development, blending established cloud-native practices with emerging AI-native approaches.

I. The Evolution of Software Delivery

Edith Harbaugh, CEO of LaunchDarkly, discusses the significant shift in software delivery practices over the past decade. In 2014, the prevailing mindset was infrequent, quarterly releases, often hampered by on-premise infrastructure and a perceived need for stability through minimal change. Gartner even advocated for “bi-modal development,” suggesting banks should minimize releases. LaunchDarkly was founded in 2014 with the mission to “launch, measure, and control software,” a mission that remains relevant today. The advent of cloud computing, particularly AWS, fundamentally altered this landscape, making infrastructure readily available and enabling more frequent deployments. Now, releasing once a quarter is considered slow, and cloud adoption is the norm.

II. The Impact of AI on Development & Delivery

The current transition, described as the “messy middle” or a “Renaissance” period, is driven by the rapid emergence of AI. While AI simplifies coding, making it easier to produce code, it introduces new challenges related to code quality. As one advisor put it, “it’s much easier to code, it’s still hard to get good code.” AI’s ease of provisioning, particularly with GPUs, presents choices and complexities. However, LaunchDarkly views AI as an extension of its core mission – helping users launch and control software. AI is essentially another “backend” that can be tested and managed using existing progressive delivery techniques.

III. Progressive Delivery: A Spectrum of Approaches

Progressive delivery, popularized by James Governor, is the practice of releasing software frequently and incrementally. LaunchDarkly’s 5,000+ customers demonstrate a wide range of approaches. Regulated industries like banking may still release monthly but utilize feature flags to queue up tested code for rapid deployment upon regulatory approval. Consumer-facing companies, conversely, may release dozens of times a day, using progressive delivery to roll out features to beta customers, specific regions, and monitor performance before wider release. This flexibility is a key benefit of the modern software delivery landscape.

IV. LaunchDarkly’s Role in AI Workload Management

LaunchDarkly is adapting to the AI era with products like “AI Configs,” which allows customers to A/B test different Large Language Models (LLMs) against their business metrics in a controlled environment. The core principle remains consistent: measure, control, and iterate. The company also offers “guarded releases,” which automatically roll back features if pre-defined metrics are breached, ensuring stability. Harbaugh emphasizes that rollbacks are essential and will remain so, regardless of the underlying technology. She highlights the ability to manage multiple versions of a feature ("old, new, new") for granular control and rollback options.

V. Data Modeling and Versioning in the AI Era

A key distinction between code versioning and data modeling versioning is the complexity of data ingestion. When data models change in a live environment, maintaining data consistency across both old and new versions is crucial to prevent data loss. This requires replication and careful cutover strategies. As AI workloads increase in complexity, dependencies and data interchange become more challenging, requiring robust versioning and management practices.

VI. The Future of Software Development: Self-Healing and Automation

Harbaugh envisions a future of self-healing and self-optimizing software. The next step beyond automated rollbacks is automated code generation based on detected errors. For example, if an issue is identified on a specific Android version, an AI agent could automatically generate a fix, check it in, and roll it out. Similarly, analyzing support tickets and automatically generating code to address common issues is a promising area. This automation builds upon existing DevOps practices and leverages AI to streamline the development lifecycle. She draws parallels to the shift from assembly language to modern languages and the impact of spreadsheets on Wall Street, suggesting that AI will not reduce the need for engineers but rather augment their capabilities.

VII. Key Takeaways & LaunchDarkly’s Vision

The core message is that LaunchDarkly’s mission – empowering developers to move faster with less risk – remains central. The increasing volume of code, driven by AI, necessitates tools and processes that enable rapid iteration and controlled experimentation. The company aims to continue providing those tools, focusing on feature-level optimization and automating responses to issues identified through monitoring and feedback. The future of software development is about leveraging AI to automate tedious tasks, allowing engineers to focus on creativity and innovation.

Notable Quote:

• “Software used to be much harder… with AI it’s much easier to code than it ever was before. But it’s still hard to get good code.” – Edith Harbaugh
• “If you tell somebody today, ‘My goal is to release once a quarter,’ like, what are you doing with all your time?” – Edith Harbaugh
• “Rollback never went away, and it never will go away. I mean, don’t you ever want to undo something?” – Edith Harbaugh

Technical Terms:

• LLM (Large Language Model): A type of AI model trained on massive datasets of text, capable of generating human-like text.
• AB Testing: A method of comparing two versions of a feature to determine which performs better.
• Microservices: An architectural style that structures an application as a collection of loosely coupled services.
• Progressive Delivery: A software release strategy focused on frequent, small releases with continuous monitoring and control.
• Feature Flags: Techniques to enable or disable functionality without deploying new code.
• Guard Rails: Safety mechanisms or constraints implemented to prevent undesirable outcomes.
• Non-Deterministic Results: Outcomes that are not predictable or repeatable, often associated with AI models.