Patrick Debois, founder of Tessl and inventor of the term DevOps, proposes in this foundational article the concept of the Context Development Lifecycle (CDLC) — an engineering cycle that treats the context of AI coding agents as a full-fledged engineering artifact, not a forgotten markdown file.
The starting observation is a fundamental shift in the bottleneck. For decades, the industry optimized how humans write code (waterfall, agile, DevOps, platform engineering). But coding agents now write the code. The new bottleneck is context quality: the structured knowledge that guides agents toward the right solutions. Each agent session is comparable to a new hire starting from scratch, with no institutional memory.
The CDLC breaks down into four phases. The "Generate" phase consists of making implicit knowledge explicit at three levels — technical (standards, architectural patterns), project (priorities, scope), and business (compliance, client expectations). Debois insists: context rots and conflicts, requiring active management of its freshness and consistency.
The "Evaluate" phase applies TDD principles to context. Scenarios are defined, agent outputs are evaluated statistically (since LLMs are non-deterministic), and each evaluation failure reveals an unwritten specification rather than an agent flaw. Evaluations also allow navigating between models factually rather than intuitively.
The "Distribute" phase treats context as a versioned, secured package. The key insight is incentive alignment: for the first time, sharing knowledge directly serves the author's own interest, since better context improves their own interactions with agents.
The "Observe" phase closes the loop. In production, agents reveal gaps through their questions, their unexpected choices, and their improvisation when context is silent. These signals feed the next cycle.
Debois draws an explicit parallel with DevOps, which he co-founded in 2009. The dynamic is similar — integrating two previously separate activities — but the structural advantage is greater: the incentive to collaborate is naturally aligned with individual interest. He concludes by warning that infinite context windows will not eliminate the need for the CDLC, since they transform the curation challenge into a governance challenge, potentially a more complex one.