Anthropic has offered a judgment worth taking seriously: the key to an Agent is not merely how well its prompts are written, but what the model actually sees at each turn of reasoning. As context keeps expanding, information recall and long-horizon reasoning can deteriorate.
What this is
Context engineering means organizing and maintaining the information currently visible to the model, including system instructions, tools, external data, and conversation history. Anthropic sees prompt engineering as focusing on “how to say it,” while context engineering focuses on “how to select it.” The goal is to keep only the information most likely to advance the task within a limited context window.
This approach is not simply about shortening context. System instructions should provide sufficiently specific direction; the toolset should be small and clear; and examples should cover typical cases. For long-running tasks, three methods can help: compressing conversations by summarizing key conclusions; using structured notes that store important information outside the context and load it when needed; and retrieving information on demand by initially retaining file paths or query entry points, then calling tools to fetch details as necessary.
Industry view
We believe this judgment is more useful for complex Agents than simply pursuing larger model context windows. Especially in tasks that run for tens of minutes, such as code migration and financial analysis, enterprises need to invest in information filtering, tool design, and process recovery capabilities—not just procure more compute.
But the risks are equally clear. On-demand retrieval increases runtime, compression may mistakenly discard details that only become important later, and more tools make it easier for the model to choose incorrectly. More importantly, Anthropic's recommendations largely come from its own products and research practices and may not transfer directly to every model or business. High-signal context requires continuous evaluation; if the filtering rules themselves are flawed, the system will merely proceed more efficiently in the wrong direction.
Impact on regular people
For enterprise IT: When enterprises deploy AI Agents, their procurement lists should include more than models and compute. They also need to build access controls, tool interfaces, record retention, and task recovery mechanisms, which are becoming the bulk of project costs.
For professionals: Long-running conversations with AI will place greater emphasis on periodic organization and capturing key conclusions. Regularly asking an assistant to summarize decisions, outstanding tasks, and supporting evidence is more likely to produce stable results than continually pasting in every piece of material.
For the consumer market: Assistants capable of handling long-running tasks reliably will gradually enter office, programming, and information-service scenarios. Reliability, however, remains a purchasing threshold. A model's ability to “remember” does not mean it “remembers correctly”; users still need to verify key figures and execution results.