Why Context Quality Is the New Frontier in AI Agent Development

DevOps has long borrowed concepts from software development to improve operations. Now, one key insight is emerging: the information and instructions fed to AI agents—what practitioners call "context"—should be managed with the same discipline applied to code itself.

At a recent tech talk, Patrick Debois, a pioneer of the DevOps movement, outlined a framework for treating context as a first-class artifact. Just as modern development teams build, test, and deploy code through structured lifecycles, they should do the same for the prompts, rules, guidelines, and documentation that tell AI agents what to do.

The Four Stages of Context Management#

Debois proposes a parallel between the software development lifecycle (SDLC) and what he calls the "Context Development Lifecycle," organized around four core activities:

Generate. Teams must curate and assemble all the context an agent will need: documentation, code guidelines, task specifications, and interconnections to external systems. This goes beyond ad-hoc prompting. Emerging standards like AGENTS.md help teams reuse context fragments—much like sharing libraries of code—while tools such as Context7 and Ref automatically pull relevant documentation for dependencies.

Evaluate. Testing context quality mirrors traditional code testing. Benchmarks and evaluation frameworks allow teams to measure whether context produces the desired agent behavior. This can range from simple syntax checks (akin to linting) to real-world tests run within an actual codebase, similar to end-to-end testing in software development.

Distribute. Context should be packaged, versioned, and shared across projects using package managers and registries—just as npm or Maven work for code libraries. Teams can standardize on common guidelines and build reusable "skills" that other teams pull in as dependencies.

Observe and Improve. Feedback loops close the cycle: teams monitor how context performs in production, collect failures or unexpected behaviors, and iteratively refine the context to improve agent outcomes.

Why Testing Context Matters#

One critical insight is that context evaluation remains largely underdeveloped. While benchmarks like SWE-bench exist for model performance, most organizations lack systematic ways to verify that their context actually works. Debois highlights the challenge: testing is non-deterministic—running the same agent multiple times may yield different results. Teams need to decide on acceptable error budgets, similar to how DevOps teams manage service level objectives.