Your agent context needs a development lifecycle

Patrick has coined a framework called the Context Development Lifecycle (CDLC) to address the lack of a proper software development lifecycle for coding agents. The CDLC consists of four phases: Generate, Evaluate, Distribute, and Observe.

During the Generate phase, teams write skills, build prompt configurations, and set up agent rules - essentially the equivalent of writing code. However, this phase often lacks proper testing. In the Evaluate phase, teams test their skills by running scenarios, checking outputs across different models, and verifying that the skill activates on the correct trigger words. This phase is crucial for testing the quality of the context being provided to the model.

The Distribute phase involves shipping skills to a repository or an installable registry with versioning, discoverability, and access controls. This is the step where skills are typically copied and pasted across teams without versioning, which can lead to inconsistencies and issues. The final Observe phase focuses on monitoring the performance of the skills in production.

This includes checking if the skills are being used, if they produce the expected results, and identifying when developers need to intervene or correct the agent's output.

The CDLC aims to treat context, which is now considered the new code, with the same level of care and rigor as software development practices. By implementing this framework, organizations can improve the quality of the context provided to their coding agents, reduce human touchpoints, and increase the reuse of skills. This, in turn, leads to better autonomous agentic coding loops and improved cost efficiency.