# debois-tessl-context-development-lifecycle-ai-coding-agents-2026-02-19

## Veille

Context Development Lifecycle (CDLC) — a context engineering cycle for AI coding agents

## Titre Article

The Context Development Lifecycle: Optimizing Context for AI Coding Agents

## Date

2026-02-19

## URL

https://tessl.io/blog/context-development-lifecycle-better-context-for-ai-coding-agents/

## Keywords

context development lifecycle, CDLC, organizational context, coding agents, evaluations, context distribution, context governance, DevOps parallel, infinite context windows, implicit knowledge, TDD for context

## Authors

Patrick Debois

## Ton

**Profile**: Founder-practitioner perspective, professional and pedagogical register, intermediate technical level. Patrick Debois, inventor of the term DevOps, writes as an architect of a movement laying the foundations of a new paradigm. **Style**: Structured, didactic prose, methodical progression from observation ("the bottleneck has shifted") to solution (the four CDLC phases). Powerful historical analogies — context before the CDLC compared to code before version control. Use of rhetorical questions to engage the reader ("You wouldn't ship code without tests. Why would you ship context without evaluations?"). Concrete metaphors: "Every session is a new hire," "context rots," "context as a package." The parallel with DevOps is central and deliberate — Debois speaks from his own history. A tone of benevolent authority aimed at tech leads and engineering managers.

## Pense-betes

- **Central thesis**: the bottleneck of software development is no longer the speed of writing code, but the quality of the context provided to agents
- **CDLC in 4 phases**: Generate (make the implicit explicit), Evaluate (TDD for context), Distribute (context as a versioned package), Observe (production feedback loop)
- **"Every session is a new hire"** — agents start from scratch every time, with no muscle memory or hallway conversations
- **Three levels of context**: technical (standards, patterns), project (timelines, scope), business (compliance, client expectations) — most teams only provide the first
- **Context rots and conflicts** — its freshness and consistency must be managed, not just its existence
- **Evaluations (evals)**: a TDD approach — define expected behavior, run, observe failure, improve context. Statistical quality control since LLMs are non-deterministic
- **Evaluation failure = unwritten specification** — the problem is rarely the agent, more often the incomplete context
- **Distribution**: developers want to share context because it directly improves their own agents — "for the first time, there's a selfish reason to share knowledge"
- **DevOps parallel**: same collaboration dynamic, but the incentive is structurally better aligned — context benefits the author AND other teams as a side effect
- **Infinite context windows do not solve the problem**: more context = more contradictions. The challenge shifts from curation to governance
- **Context as an attack surface** — requires the same supply-chain security posture as dependencies
- **Precursor article** to the "Context Flywheel" concept, which develops the compounding effect over time

## RésuméDe400mots

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.

## GrapheDeConnaissance

- Patrick Debois —a_créé→ Context Development Lifecycle (METHODOLOGIE, 0.97)
- Patrick Debois —a_créé→ Tessl (ORGANISATION, 0.9)
- Patrick Debois —a_créé→ DevOps (CONCEPT, 0.85)
- phase Générer —fait_partie_de→ Context Development Lifecycle (METHODOLOGIE, 0.98)
- phase Évaluer —fait_partie_de→ Context Development Lifecycle (METHODOLOGIE, 0.98)
- phase Distribuer —fait_partie_de→ Context Development Lifecycle (METHODOLOGIE, 0.98)
- phase Observer —fait_partie_de→ Context Development Lifecycle (METHODOLOGIE, 0.98)
- CDLC —s_inspire_de→ DevOps (METHODOLOGIE, 0.92)
- Phase Évaluer —est_basé_sur→ Test-Driven Development (METHODOLOGIE, 0.93)
- Phase Distribuer —utilise→ contexte comme package versionné (CONCEPT, 0.9)
- Patrick Debois —affirme_que→ le contexte est le nouveau goulot d'étranglement (AFFIRMATION, 0.95)
- Agents de codage —utilise→ contexte structuré (CONCEPT, 0.96)
- Patrick Debois —affirme_que→ les fenêtres de contexte infinies ne résolvent pas la gouvernance du contexte (AFFIRMATION, 0.88)
- Patrick Debois —affirme_que→ l'incitation au partage est naturellement alignée (AFFIRMATION, 0.9)
- Patrick Debois —affirme_que→ le contexte pourrit et entre en conflit (AFFIRMATION, 0.92)

---
Canonical: https://www.thekb.eu/en/fiches/debois-tessl-context-development-lifecycle-ai-coding-agents-2026-02-19/