[KongchangAI]
ConceptAI Agent / AI代理 / Agentic AI

Agent

An Agent, or intelligent agent, is a program or system in artificial intelligence that can perceive its environment, make autonomous decisions, and take actions to accomplish goals. Its core characteristics include autonomy—operating without continuous human intervention—goal-directedness, enabling the planning and execution of multi-step tasks, and tool-calling capability, allowing interaction with external systems or APIs. Agents are widely used in task automation, dialogue systems, and programming assistance.

Core Facts

Timeline (last 90 days)

Nov 17

After an Agent performs an operation, the state in the system of record is the final basis of fact, whereas tracing logs record the Agent's intent and the tool responses.

Unverified50%
Nov 16

Most continual learning work focuses only on changes in data distribution, whereas in agent scenarios the capability contract itself also changes, and the model needs to adapt simultaneously to changes in the input distribution and the available action space.

Unverified50%
Nov 16

In low-frequency but high-risk scenarios such as power outages, an agent whose behavior is predictable but whose capabilities are limited may be more suitable for dispatch duties than a more capable but unpredictable agent.

Unverified50%
Nov 16

Before deploying agents in homes at scale, resilience verification for outage scenarios should become a standard step rather than remedial work afterward.

Unverified50%
Nov 15

Sean summarizes the evolution of Agents in four stages: a smart chat Agent, an Agent that can use tools and retrieve memory, an always-on personal assistant, and a true teammate.

Unverified50%
Nov 15

If the next model significantly improves long-context understanding or function-calling reliability, existing agent planning modules, memory modules, and error-recovery mechanisms may all need to be rebuilt.

Unverified50%
Nov 14

The solution proposed in the video is that in agent scenarios, the ideal state is for the agent itself to decide "I want to use this skill," rather than leaving all decisions to the user.

Unverified50%
Nov 14

Agent-type products usually need to hold OAuth tokens or API keys (email, calendar, payments), which amounts to storing long-term credentials for multiple accounts together in a single process, creating a single point of compromise risk.

Unverified50%
Nov 14

Vulnerabilities in ordinary applications may leak data, whereas vulnerabilities in an Agent may directly cause financial loss or send erroneous messages in the user's name.

Unverified50%
Nov 14

In personal Agent scenarios, least privilege, short-lived tokens and secondary confirmation for high-risk operations are not optional but essential.

Unverified50%

40 more timeline events

All Facts (20)

Verified

Adopting a human-in-the-loop collaborative mode to introduce human checkpoints can avoid the "hallucination cascade" of error accumulation in purely automated generation, forming an iterative loop of generate–review–correct–continue.

90%
Verified

A sandbox is a technical mechanism that restricts program execution to an isolated environment, with the core idea that even if the isolated program exhibits malicious behavior, it cannot affect the host system or external environment.

90%
Verified

Mainstream Agent architectures typically use a large language model as the brain, combined with memory modules, planning modules, and tool interfaces to accomplish complex tasks

90%
Verified

Agents possess capabilities for planning, tool invocation, memory, and autonomous iteration, distinguishing them from the single-round question-and-answer mode of traditional large language models.

90%
Verified

A Skill is a modular encapsulation that packages specialized capabilities, domain knowledge, and operational norms into reusable units, enabling Agents to invoke them automatically at the appropriate moment.

90%
Verified

Autonomous Agents perform a 'perception-planning-action-observation' loop based on the ReAct (Reasoning + Acting) framework and have higher requirements for the quality of the initial context.

90%
Verified

The "perception-decision-action" loop is the technical foundation of the ReAct (Reasoning + Acting) paradigm.

90%
Verified

A complex Agent task may require dozens or even hundreds of model calls

90%
Verified

Since 2023, with the leap in reasoning capabilities of large language models such as GPT-4 and Claude, Agents have rapidly moved from academic concepts to engineering practice.

90%
Verified

LangGraph abstracts agent workflows as directed graphs, where each node represents an agent or processing step, and edges represent state transition conditions.

90%
Verified

Agent planning capabilities are typically implemented through the ReAct (Reasoning + Acting) framework.

90%
Verified

In the ReAct framework, when the number of tools exceeds 20, both token consumption and the model's tool selection accuracy drop significantly, a problem known as 'tool overload'.

90%
Verified

AI-Native products center on intent expression, where users describe their goals and the tool is responsible for finding the execution path; the industry also refers to this as a Conversation-First architecture.

85%
Verified

The classic paradigm of Agent workflows is the separation of the planning layer and the execution layer, where the planning layer is responsible for understanding goals, decomposing tasks, and generating execution plans, while the execution layer handles actual operations via dedicated tools or smaller models.

85%
Verified

A typical Agent is composed of a cycle of 'Perception—Planning—Action—Reflection'.

85%
Verified

An agent's behavioral chain is essentially a probabilistic reasoning process in which the same input may produce different tool-call sequences, which is what makes agent evaluation difficult.

85%
Verified

Agent latency and cost grow linearly or even super-linearly with the number of tool calls.

80%
Verified

The more thorough the sandbox isolation, the more constrained the Agent's capabilities; the more open the permissions, the harder the risks are to control.

80%
Verified

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%
Verified

An agent's long-term memory can be subdivided into three categories: episodic memory, semantic memory, and procedural memory.

80%

Source Articles