Presentation: Context Is the New Code

Patrick Debois, a practitioner and researcher exploring AI agents' impact on software development, argues that context is the new code. He suggests treating context like code by incorporating testing, CI/CD, package managers, and security scanning. This approach enables engineering leaders to scale AI coding agents reliably, maintain control over non-deterministic outputs, and build long-term organizational knowledge.

Debois shares two examples demonstrating how context functions like code. The first involves Andrej Karpathy's view that AI agents generate code through prompts and context. The second example comes from Debois' experience at his company. They faced numerous coding challenges during onboarding, eventually realizing that conveying the required scenarios in natural language and providing them to the AI agent simplified the process. Debois likens this to compressing multiple lines of code into just a few words.

Debois proposes a parallel between the Software Development Life Cycle and the Context Development Life Cycle. He outlines the four steps: generate, evaluate, distribute, and observe. Generate refers to the human context generator, such as typing into a Claude Code or prompting an AI agent. Evaluate involves determining whether the provided context is suitable and distributing it to team members, the organization, and AI agents.

Observe entails assessing the effectiveness of the context and making necessary improvements or generating new context, forming a loop similar to the Software Development Life Cycle.

The presentation covers four aspects of context: generation, evaluation, distribution, and observation. Generation involves humans creating context through prompts and instructions. Evaluation assesses the quality of the context, while distribution transmits it to various stakeholders and agents. Observe analyzes the results and facilitates iterative improvement.

Debois also discusses various types of context, including CLAUDE.md, code and documentation, libraries, connectors, specifications, and intent integrity. He emphasizes the importance of standardization, citing AGENTS.md as a potential industry-wide standard. The presentation concludes with the concept of the context flywheel, illustrating how generating, evaluating, distributing, and observing context create a continuous loop akin to the traditional Software Development Life Cycle.