Context engineering supersedes prompt engineering for agents¶
Insight: Context engineering is replacing prompt engineering as the key skill for building effective AI agents. A typical Manus agent task requires ~50 tool calls, making context management — not just prompting — the critical challenge. The "bitter lesson" in AI engineering is that scaling and letting models do more with better context beats elaborate hand-crafted pipelines. Memory management, human-in-the-loop patterns, and MCP-based tool integration are the practical surfaces where context engineering plays out.
Detail: Lance Martin's presentation covers: (1) Context grows exponentially with agents — managing what goes in and out of the context window is the engineering challenge; (2) The bitter lesson applied to AI engineering means investing in context infrastructure rather than clever prompts; (3) Framework standardization (via MCP for protocols, orchestration tools for workflows) is the practical path forward; (4) RAG benchmarking and retrieval quality directly affect agent performance through context quality.
Sources
Related: existing entry "MCP has fundamental security and UX problems" in external/mcp.md — COMPLEMENTS