Patrick Debois founded DevOpsDays, helped give DevOps its name, and co-authored The DevOps Handbook. Now the Product DevRel lead at Tessl, he applies the lessons of software automation, operational reliability, and organizational change to AI-native development and the emerging discipline of engineering the context that guides coding agents.
Debois organized the first DevOpsDays in Ghent in 2009, led the conference community through 2014, and remains an advisory organizer. His subsequent career has included engineering leadership and work with Atlassian and Snyk. His open-source projects include Veewee, which simplifies building Vagrant boxes, and Sahara, a Vagrant plugin for managing sandbox states.
As generative AI entered software development, Debois created open learning materials covering LLMs, generative AI, and DevSecOps and became a curator of the AI Native Developer community. At Tessl, his work centers on making agent instructions, reusable skills, evaluations, and accumulated organizational knowledge reliable enough for real engineering teams.
AI platform engineering: Debois argues that organizations adopting generative AI need shared capabilities for model access, data connections, evaluation, observability, security, and developer enablement. His platform-engineering approach combines centralized infrastructure and baseline governance with product-team experimentation and application-specific safeguards. He warns that faster code generation can increase review burdens and weaken engineers’ situational awareness.
Four shifts in AI-native work: His framework for AI-native development describes developers moving from producing code to supervising agents, from implementation to specifying intent, from delivery to product discovery, and from generating content to preserving organizational knowledge. Agent permissions, specifications, prototypes, onboarding materials, incident lessons, and rejected feature decisions all become inputs to better future work.
Context Development Lifecycle: Debois organizes agent-context engineering into four stages: generate, evaluate, distribute, and observe. In his account of context as engineered infrastructure, AGENTS.md and CLAUDE.md files, current documentation, specifications, reusable skills, and MCP-connected information require versioning, validation, security scrutiny, and operational feedback. Shared skill registries introduce familiar challenges around dependencies, provenance, conflicting instructions, and unsafe third-party content.
Evaluation error budgets: Because identical agent evaluations can produce different results, Debois proposes repeated trials and failure tolerances calibrated to actual risk. His writing on CI/CD for context distinguishes consequential security or architectural requirements from lower-stakes preferences and treats production failures as opportunities to improve both evaluations and instructions.
The context flywheel: Debois sees organizational knowledge as a compounding advantage: agent logs, pull-request feedback, and production incidents reveal missing context; improved instructions then benefit other developers and teams. In a public explanation of the context flywheel, he argues that context quality can differentiate organizations as coding models and tools become more interchangeable. His exploration of Reverse Conway’s Law extends that question to how agents might reshape team boundaries, specialist roles, onboarding, and management, while retaining human judgment and preparedness for failure.
Scaling GenAI across an organization requires shared infrastructure, practical enablement, and governance that helps application teams build and operate useful products.
Reusable instructions need the same engineering care as software: evaluation, distribution, dependency management and feedback from the agents and teams that use them.
As agents move from personal copilots to organizational peers, the design problem shifts from where AI belongs to who specifies, reviews, governs, and benefits from its work.
As agents take on implementation, developers spend more effort reviewing work, specifying intent, discovering useful products, and preserving what the team learns.
Patrick Debois explains why coding-agent adoption becomes an organizational systems problem: teams must turn repeated corrections into reusable context, maintained harnesses, supported platform paths, and risk-adjusted autonomy.
Treat repeated agent corrections as signals to improve shared context, harnesses, and loops—not merely as code-review chores.
