Agent
Core Facts
AI evaluation targets are divided into three layers—large language model evaluation, knowledge base (RAG) evaluation, and agent (Agent) evaluation—forming a progressive relationship of increasing complexity and knowledge depth.
AI智能体具备三个关键能力：任务规划、工具调用、记忆与上下文管理
AI Agent是一种能够自主感知环境、做出决策并执行操作的人工智能系统，与传统AI聊天机器人不同，它具有目标导向性、自主规划能力和调用外部工具的能力
Timeline (last 90 days)
Agent behavior is jointly determined by system prompts, tool descriptions, and a small number of examples, and changes in business rules often require synchronous modification of multiple prompts.
The reliability issues of agents have given rise to "human-in-the-loop" authorization designs, pre-operation confirmation mechanisms and sandboxed execution environments.
Agents often confidently claim the problem is somewhere, but checking the actual tool-call log reveals it never read the affected code; this kind of hallucination is where trust begins to collapse.
Agents communicating through the CPU cache can avoid network stack overhead, theoretically achieving nanosecond-level latency.
Agents often hold full credentials, injected instructions can trigger actual external operations, and there is no human approval step throughout the process.
The core point in the Reddit discussion is that agents "cannot distinguish legitimate orchestration from injected content."
The stronger an agent's capabilities and the greater its permissions, the broader the potential damage surface of injection attacks; an assistant that can only read text and an agent that can call payment APIs and send emails do not face the same injection vector at the same order of risk.
The agent has no reliable mechanism to distinguish between "document content provided by the user" and "execution instructions issued by the system"; it inherits the ambient trust of its runtime environment.
Permission checks at the tool-call layer are more critical, and agents should be forced to require human confirmation or secondary authorization before performing sensitive operations, especially in scenarios involving write operations, external communication, and credential use.
When vendors emphasize improvements in agentic capabilities, they usually need to also explain permission boundaries, sandbox mechanisms and abuse protection strategies.
40 more timeline events
All Facts (20)
AI evaluation targets are divided into three layers—large language model evaluation, knowledge base (RAG) evaluation, and agent (Agent) evaluation—forming a progressive relationship of increasing complexity and knowledge depth.
80%VerifiedAI智能体具备三个关键能力：任务规划、工具调用、记忆与上下文管理
80%VerifiedAI Agent是一种能够自主感知环境、做出决策并执行操作的人工智能系统，与传统AI聊天机器人不同，它具有目标导向性、自主规划能力和调用外部工具的能力
80%VerifiedWhat sets AI apart from previous technologies is its unprecedented pervasiveness and autonomy—it can autonomously generate decisions, produce content, and drive behavior.
80%VerifiedAn agent is a system built around a generative model that can plan steps toward a goal, call external tools, observe results, and iteratively adjust.
75%VerifiedThe technical implementation of persistent agents typically relies on core components such as persistent memory, tool-calling capabilities, and task-scheduling mechanisms.
75%VerifiedAgent latency and cost grow linearly or even super-linearly with the number of tool calls.
75%VerifiedTasks that are mechanically repetitive and time-consuming should be delegated to AI, while tasks requiring deep understanding, shaping judgment, and building core competencies should be handled personally.
75%VerifiedThe essence of an agent is to equip a model with the capabilities of planning, tool calling, multi-step reasoning, and environment interaction.
75%VerifiedRAG（检索增强生成）是一种结合信息检索与生成式大模型的架构范式，模型生成答案前先根据用户查询从外部知识库检索最相关的文档片段，再将其作为上下文注入模型
70%VerifiedThe agent loop follows a 'reasoning-action-observation' paradigm, iterating repeatedly until the goal is achieved or a stopping condition is reached.
70%VerifiedThe classic industry formula is: Agent = LLM + Planning + Memory + Tools.
70%VerifiedThe article's author believes that the core competitiveness in the agent era lies in whether one can deeply integrate professional experience with AI capabilities, rather than how many tools one knows how to use.
70%VerifiedThe core mechanism of an agent is: the model issues a tool call request → code runs the tool → the execution result is returned to the model → the model decides the next step, forming a closed loop.
65%VerifiedThe model itself never runs code; it only writes out requests to call tools, and it is external code that actually executes the tools
65%VerifiedEmbodied intelligence refers to the research direction of embedding intelligent systems into robots or agents with physical bodies that can directly interact with the real environment.
65%VerifiedIntelligent agent products can be categorized into three types: specialized agents, general-purpose agents, and platform-building agents.
65%VerifiedWhether an agent requirement is viable can be judged based on whether the task structure is clear, whether the data is controllable, and whether the results are verifiable.
65%VerifiedHigh-level agents introduce ReAct (Reasoning + Acting) or multi-step planning capabilities, enabling them to decompose goals autonomously, dynamically select tools, and adjust subsequent steps based on intermediate results.
65%VerifiedAn AI system involves multiple parties including data providers, model trainers, fine-tuners, deployment platforms, and end users, making the chain of responsibility more complex than that of a firearm.
65%