Anthropic's 'Context Engineering' Boosts AI Agents' Long-Term Coherence
Anthropic, the AI company behind the Claude model, has introduced a 'gemini' method called 'context engineering' to help AI agents manage their limited attention and maintain coherence in long-term tasks. This shift comes as LLMs face 'context decay' due to architectural constraints. The company has also released a memory tool in public beta as part of the Claude 4.5 Sonnet update.
Anthropic has observed a trend towards 'just in time' strategies, where agents retain light identifiers and load data dynamically at runtime. This approach is evident in Anthropic's Claude Code for complex data analyses. However, LLMs have a limited 'attention budget', leading to 'context decay' where the model's ability to recall information decreases with increasing token numbers.
Anthropic's 'gemini' engineering considers the entire context state, including system instructions, tools, external data, and message history, unlike prompt engineering which focuses mainly on writing effective prompts. The company recommends finding the right 'height level' for system prompts to balance behavior control and flexibility for strong heuristics, and to use token-efficient tools with minimal functional overlaps.
The new memory tool allows agents to build knowledge databases over time, reducing token consumption. In a 100-round web search, this combination increased agent search performance by 39%, with context editing alone bringing a 29% improvement, and an 84% decrease in token consumption. For multi-step tasks, Anthropic uses three main techniques: compaction, structured notes, and sub-agent architectures to manage context effectively. The new features are now available in public beta on the Claude Developer Platform, also via Amazon Bedrock and Google Cloud Vertex AI, with documentations and a cookbook provided.
Anthropic's 'gemini' engineering and memory tool address the challenge of 'context decay' in LLMs, enabling agents to maintain coherence in long-term tasks and improving search performance. These innovations are now available in public beta, offering AI developers new tools to enhance their agents' capabilities.
