Latest

Anthropic Says Context Engineering Goes Beyond Prompt Writing for Multi-Turn Agents

According to Anthropic, building capable multi-turn agents requires managing the entire context state—system instructions, tools, MCP connections, external data, and message history—not just crafting effective prompts.

What context engineering actually covers

Prompt engineering focuses on how to write effective instructions—primarily system prompts. Context engineering, as described by Anthropic in a September 2025 post, extends that concern to every token the model sees at inference time: system instructions, tools, Model Context Protocol (MCP) connections, external data, and message history. The shift matters because an agent running across multiple turns of inference accumulates data continuously, and decisions about what to keep, compress, or discard happen at every cycle—not just at setup.

Anthropic also notes a phenomenon they call context rot: as token count in the context window grows, a model's ability to accurately recall information from that context decreases. This degradation appears across models, though its steepness varies. The practical implication is that adding more information does not straightforwardly improve agent behavior—curation is load-bearing work, not a nice-to-have.

What to try now, what to verify first

The Anthropic framing gives a clear starting point for teams building multi-turn agents: audit what lands in the context window at each inference step, not just what appears in the system prompt. Treating context as a resource to be curated—rather than a container to be filled—is a design posture that follows directly from the documented scope in Anthropic's engineering post.

Several questions are worth holding open before acting on the vendor framing: How severe is context rot on the specific model and task you are using? Does MCP retrieval in practice introduce latency or inconsistency that affects curation decisions? What compression or summarization strategies work for your domain's message history? Chooscope has not tested these independently. The post covers Anthropic's observations and positions; gains in specific deployments remain for teams to measure.

Sources

  1. Anthropic — 2026-10-18

Change history

  • rev.1Published

Display language

Related