Context is what the engineer deliberately puts in front of the agent, rather than whatever the agent happens to encounter. Context engineering is a discipline with its own principles, artefacts, and failure modes.

Context as a budgeted resource: the token window is finite, attention quality degrades toward the end of large contexts, and kitchen-sink context (providing everything that might be relevant) demonstrably degrades performance. The engineering task is curation.

Context as a designed artefact: the CLAUDE.md and AGENTS.md files are the primary mechanisms for persisting project-specific knowledge across sessions. Recent empirical research makes the design requirement concrete: human-written context files improve agent performance by approximately 4%, while LLM-generated context files decrease performance by 3% and increase token cost by over 20%.

What goes into such a file falls into four kinds of content. Standards and processes: coding conventions, naming rules, build steps, the testing approach. Code quality expectations: stylistic guidance, the team’s stance on readability versus performance. Problem-solving heuristics: how the team prefers tasks to be approached. Collaboration protocols: how the agent should communicate, what to flag for human review, when to escalate.

A related practice is a maintained lessons-learned file: a running record of corrections the human has made and rules the agent should internalise across tasks. This converts ephemeral course-corrections into durable project knowledge. Context as a maintained system: context files change as the project changes. Bad context is the dominant cause of bad agent output over time, and it accumulates silently.