Context Is the New Code — Patrick Debois, Tessl

Patrick Debois, known as the father of DevOps and now working at Tessl, presents a framework he calls the Context Development Lifecycle (CDL) — a structured engineering discipline for managing the prompts, rules, memory, and instructions that drive AI coding agents. Drawing an explicit parallel to his 2009 DevOps insight ("what if ops looked more like dev?"), Debois argues that context is undergoing the same transformation code once did, and that teams need the same rigour applied to context that they apply to software: versioning, testing, packaging, and observability.

Debois opens by asserting that as AI coding agents become mainstream, context — not code — is the primary creative input. He notes that large procedural code helpers are being collapsed into reusable agent "skills" described in natural language, and that even entire onboarding workflows (detecting package managers, ecosystems, then walking users through steps) that would require hundreds of lines of code can now be expressed as a brief skill definition. He frames this as code transforming back into context.

The CDL consists of four phases — Generate, Evaluate, Distribute, and Observe — arranged as a continuous improvement loop.

**Generate** covers the full range of context creation: ad hoc prompting, reusable instruction files (agent.md / CLAUDE.md), pulling in live library documentation to prevent hallucinated API versions, MCP-sourced context from GitHub, GitLab, Slack, and ticketing systems, and spec-driven development where a specification prompt is decomposed by the agent into a planning-mode step-by-step execution plan.

**Evaluate** is the phase Debois considers most under-adopted. He introduces a multi-tier testing model for context that mirrors software testing: linting (structural validation of skill definitions — e.g., enforcing description length limits), a "Grammarly" layer (asking an LLM whether the written context is actually intelligible and complete), unit-test-style LLM-as-judge checks (e.g., verifying generated code contains a company-mandated URL prefix like "/awesome/user"), and end-to-end tests where a judge agent is given tools and a sandbox to actually execute the generated code. Because LLM evaluation is non-deterministic, he recommends running evals multiple times (e.g., 5 runs) and using error budgets rather than binary pass/fail gates. He also covers CI/CD integration of evals and automated context optimization — using eval feedback to instruct the LLM to rewrite and improve the context itself.

**Distribute** introduces context packaging analogous to software package management: context libraries, registries (such as Tessl's own skills marketplace), versioned context packages with dependency resolution, and context security scanning (Snyk for credential leakage and prompt injection). He raises the concept of an AI SBOM (Software Bill of Materials) — tracking which model generated a skill, under what conditions, and what it depends on. He notes that while public skill registries exist, the vast majority of published skills do not pass any quality eval bar, and that enterprises are building private registries.

**Observe** closes the loop. Debois advocates mining agent logs (noting emerging standards for agent log formats) to surface patterns where agents report missing context — then creating and distributing that context organization-wide. He connects PR review feedback directly to context quality: rather than re-litigating the same PR comment, teams should improve the context so the agent doesn't repeat the mistake. He also discusses production observability: instrumenting generated code in production, capturing failures with input/output pairs, and automatically generating test cases from production errors. He warns about adversarial context scenarios — agent sandboxing is insufficient on its own because agent.md and skill files load automatically — and advocates a "context filter" analogous to a WAF (Web Application Firewall) for prompt injection and pattern filtering.

Debois concludes with a flywheel model: individual context hygiene feeds team-level context libraries, which feed cross-team registries, compounding improvement across the organization. His closing message: LLMs are just the engine — context is the fuel, and engineering the fuel is the leverage point available to every team today.