Context engineering overtakes RAG
Context engineering is starting to supplant basic RAG pipelines — an industry analysis argued that teams must build structured, evolving context (task graphs, workflow state, memories) rather than rely on raw retrieval alone. The change matters because RAG struggles with semantic drift and stale knowledge at enterprise scale, so hybrid approaches that combine structured context with retrieval are becoming the new baseline.
LangChain published "Context Engineering for Agents" on July 2, 2025 and formalized four tactics—write, select, compress, isolate—to stitch dynamic context into agent trajectories. blog.langchain.com Anthropic released an engineering guide framing context engineering as iterative curation of what enters the model's window rather than one-off prompt tweaks, published in 2025. anthropic.com Google’s ADK and a November 2025 Sessions & Memory whitepaper describe context as a “compiled view” over session state, long/short memory tiers, and orchestrated task graphs for production agents. raphaelmansuy.github.io Enterprise RAG pipelines migrated by mid‑2024 to hybrid stacks—embedding-based retrieval plus query transformation and cross‑encoder reranking—to survive production constraints, according to a practitioner guide. applied-ai.com Operational monitoring must track both data drift and concept drift in retrieval to avoid relevance decay, as shown in a retrieval‑drift monitoring playbook that separates those two failure modes. apxml.com Several implementation studies report cross‑encoder reranking can boost NDCG by ~35–45% but increase latency by roughly 5×, prompting patterns that limit rerankers to the top‑20–200 candidates or move reranking off CPUs to meet SLOs. johal.in Industry analysts and vendors now treat context engineering as the default architecture: Gartner recommended making it a strategic priority in 2025, Microsoft’s Copilot Studio documents governed RAG patterns for enterprise apps, and a market analysis projects AI/knowledge platforms to grow from $1.24B in 2024 to $67.42B by 2034. gartner.com
