Managing Context in Agentic Coding

One key point that I teach in my Agentic Coding class is the importance of managing context effectively.

Every technique in coding harnesses - be it grep, skills, subagents, CLIs and so on ultimately tie back into maintaining context relevance. Massive contexts not only confuse the agent, but also blow up the token costs. While this might have been less an issue with fixed price unlimited plans, that era is ending. With per-token API pricing becoming the norm, and price caps being implemented in organisations, we are going to see a lot of focus coming back into using the context efficiently.

My most used command is /new to wipe the context and start a new session. Yes, context auto-compacts, or you could use the /compact command manually but there is no control what the agent will decide to retain. Many times critical instructions are lost during compaction.

Instead, my preferred workflow is to externalise memory into markdown files, create new sessions liberally and rehydrate relevant memory for the current task back from the files.