Context Is the New Code: Why Your Team Needs DevOps for Agent

The proposal. Patrick argues that context—rules, docs, specs given to AI agents—should be managed like code, with its own DevOps lifecycle: generate, evaluate, distribute, observe. He draws a direct parallel to the code SDLC, calling this the "context development life cycle." The goal is systematic curation, not ad-hoc prompting.

Why now. The industry is standardizing around agent.md and skills formats, making context portable across tools. Vendors like Context7 and Ref provide curated context for libraries, and platforms like Unblocked mine Slack and docs for institutional knowledge. Patrick notes that Kuro, Spec, and Baroo treat task specifications as a new form of context. This infrastructure is nascent but forming fast.

The testing challenge. Evals are the equivalent of unit tests for context. Patrick describes linting (checking metadata syntax), AI-as-judge (Grammarly for context), and functional evals using Docker containers that run agents against test tasks. But results are non-deterministic, so teams must adopt risk-based decisions akin to error budgets. He advises capturing agent failures as new test cases to build a virtuous loop.

Distribution as packaging. Context should be versioned, packaged, and shared like libraries—not copy-pasted on Slack. Patrick envisions registries of skills with metadata, testing, and security scanning. He points to projects like G blame for provenance and a context bill-of-materials. The emerging challenge is determining ownership when context spans teams.

Observability loops. Observability for agents includes logs, self-reflection (like Hugging Face's retrospectives), and feedback from commits, Slack threads, and production incidents. Patrick demoed a prototype context firewall that intercepts file reads to detect unauthorized access, similar to a web application firewall. The goal is a feedback loop that improves context over time.

The organizational shift. Engineers now willingly write tests and documentation because it directly improves agent output—a rare alignment with good engineering practices. The speaker argues that context curation will become a differentiator for businesses, and platform teams should treat context as a first-class artifact. Ownership models, maturity levels, and chaos engineering for context remain open questions.