您的浏览器禁用了JavaScript(一种计算机语言,用以实现您与网页的交互),请解除该禁用,或者联系我们。[谷歌]:智能体伴侣 - 发现报告

智能体伴侣

2025-02-01-谷歌c***
AI智能总结
查看更多
智能体伴侣

February 2025AcknowledgementsEditors & curatorsAnant NawalgariaContent contributorsAnant NawalgariaSteven JohnsonHussain ChinoyDesignerMichael Lanning IntroductionAgent OpsAgent Success MetricsAgent EvaluationAssessing Agent CapabilitiesEvaluating Trajectory and Tool UseEvaluating the Final ResponseHuman-in-the-Loop EvaluationMore about Agent EvaluationMultiple Agents & Their EvaluationUnderstanding Multi-Agent ArchitecturesMulti-Agent Design Patterns and Their Business ImpactImportant components of AgentsChallenges in Multi-Agent systemsMulti-Agent EvaluationTable of contents 6812141517202122232425283132 Agentic RAG: A Critical Evolution in Retrieval-Augmented Generation33Agentic RAG and its Importance34Better Search, Better RAG36Agents in the enterprise38Manager of agents38Google Agentspace40NotebookLM Enterprise41Google AgentSpace Enterprise43From agents to contractors46Contracts46Contract Lifecycle49Contract execution49Contract Negotiation50Contract Feedback51Subcontracts51Automotive AI: Real World Use of Multi-Agent Architecture54Specialized Agents54Conversational Navigation Agent54Conversational Media Search Agent56Message Composition Agent56Car Manual Agent57General Knowledge Agent58Patterns in Use58 Hierarchical Pattern58Diamond Pattern60Peer-to-Peer62Collaborative Pattern64Response Mixer Agent66Adaptive Loop Pattern67Advantages of Multi-Agent Architecture for Automotive AI68Agent Builder69Summary70Endnotes74 February 2025IntroductionGenerative AI agents mark a leap forward from traditional, standalone language models,offering a dynamic approach to problem-solving and interaction. As defined in the originalAgents paper, an agent is an application engineered to achieve specific objectives byperceiving its environment and strategically acting upon it using the tools at its disposal.The fundamental principle of an agent lies in its synthesis of reasoning, logic, and access toexternal information, enabling it to perform tasks and make decisions beyond the inherentcapabilities of the underlying model. These agents possess the capacity for autonomousoperation, independently pursuing their goals and proactively determining subsequentactions, often without explicit instructions.The future of AI is agentic. February 2025The architecture of an agent is composed of three essential elements that drive its behaviorand decision-making:•Model:Within the agent's framework, the term "model" pertains to the languagemodel (LM) that functions as the central decision-making unit, employing instruction-based reasoning and logical frameworks. The model can vary from general-purpose tomultimodal or fine-tuned, depending on the agent's specific requirements.•Tools:Tools are critical for bridging the divide between the agent's internal capabilitiesand the external world, facilitating interaction with external data and services. Thesetools empower agents to access and process real-world information. Tools can includeextensions, functions, and data stores. Extensions bridge the gap between an API andan agent, enabling agents to seamlessly execute APIs. Functions are self-containedmodules of code that accomplish specific tasks. Data stores provide access to dynamicand up-to-date information, ensuring a model’s responses remain grounded in factualityand relevance.•Orchestration layer:The orchestration layer is a cyclical process that dictates how theagent assimilates information, engages in internal reasoning, and leverages that reasoningto inform its subsequent action or decision. This layer is responsible for maintainingmemory, state, reasoning, and planning. It employs prompt engineering frameworks tosteer reasoning and planning, facilitating more effective interaction with the environmentand task completion. Reasoning techniques such as ReAct, Chain-of-Thought (CoT), andTree-of-Thoughts (ToT) can be applied within this layer.Building on these foundational concepts, this companion paper is designed for developersand serves as a "102" guide to more advanced topics. It offers in-depth explorations of agentevaluation methodologies and practical applications of Google agent products for enhancingagent capabilities in solving complex, real-world problems. 7 February 20258While exploring these theoretical concepts, we'll examine how they manifest in real-worldimplementations, with a particular focus on automotive AI as a compelling case study. Theautomotive domain exemplifies the challenges and opportunities of multi-agent architecturesin production environments. Modern vehicles demand conversational interfaces that functionwith or without connectivity, balance between on-device and cloud processing for bothsafety and user experience, and seamlessly coordinate specialized capabilities acrossnavigation, media control, messaging, and vehicle systems. Through this automotive lens,we'll see how different coordination patterns -- hierarchical, collaborative, and peer-to-peer -- come together to create robust, responsive user exp