Context engineering tips for coding agents

As coding agents become a daily workflow tool, context hygiene is becoming the differentiating skill between developers who get great results and those who fight constant drift — this post nails the fundamentals.

• –Context hierarchy matters: system instructions → harness instructions (AGENTS.md) → conversation history, in that order of influence
• –Keep AGENTS.md minimal and project-specific; only codify behaviors that consistently diverge from model defaults
• –User-guided subagent delegation beats autonomous delegation for preserving context quality
• –Context compaction (auto-summarization at ~90% capacity) isn't a failure state — just ask the model to recount its understanding to restore alignment
• –The core mental model shift: context engineering is about maintaining intent alignment, not maximizing tokens